1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/fs.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <linux/wait.h>
9 #include <linux/writeback.h>
10 
11 #include "super.h"
12 #include "mds_client.h"
13 #include <linux/ceph/decode.h>
14 #include <linux/ceph/messenger.h>
15 
16 /*
17  * Capability management
18  *
19  * The Ceph metadata servers control client access to inode metadata
20  * and file data by issuing capabilities, granting clients permission
21  * to read and/or write both inode field and file data to OSDs
22  * (storage nodes).  Each capability consists of a set of bits
23  * indicating which operations are allowed.
24  *
25  * If the client holds a *_SHARED cap, the client has a coherent value
26  * that can be safely read from the cached inode.
27  *
28  * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
29  * client is allowed to change inode attributes (e.g., file size,
30  * mtime), note its dirty state in the ceph_cap, and asynchronously
31  * flush that metadata change to the MDS.
32  *
33  * In the event of a conflicting operation (perhaps by another
34  * client), the MDS will revoke the conflicting client capabilities.
35  *
36  * In order for a client to cache an inode, it must hold a capability
37  * with at least one MDS server.  When inodes are released, release
38  * notifications are batched and periodically sent en masse to the MDS
39  * cluster to release server state.
40  */
41 
42 
43 /*
44  * Generate readable cap strings for debugging output.
45  */
46 #define MAX_CAP_STR 20
47 static char cap_str[MAX_CAP_STR][40];
48 static DEFINE_SPINLOCK(cap_str_lock);
49 static int last_cap_str;
50 
gcap_string(char * s,int c)51 static char *gcap_string(char *s, int c)
52 {
53 	if (c & CEPH_CAP_GSHARED)
54 		*s++ = 's';
55 	if (c & CEPH_CAP_GEXCL)
56 		*s++ = 'x';
57 	if (c & CEPH_CAP_GCACHE)
58 		*s++ = 'c';
59 	if (c & CEPH_CAP_GRD)
60 		*s++ = 'r';
61 	if (c & CEPH_CAP_GWR)
62 		*s++ = 'w';
63 	if (c & CEPH_CAP_GBUFFER)
64 		*s++ = 'b';
65 	if (c & CEPH_CAP_GLAZYIO)
66 		*s++ = 'l';
67 	return s;
68 }
69 
ceph_cap_string(int caps)70 const char *ceph_cap_string(int caps)
71 {
72 	int i;
73 	char *s;
74 	int c;
75 
76 	spin_lock(&cap_str_lock);
77 	i = last_cap_str++;
78 	if (last_cap_str == MAX_CAP_STR)
79 		last_cap_str = 0;
80 	spin_unlock(&cap_str_lock);
81 
82 	s = cap_str[i];
83 
84 	if (caps & CEPH_CAP_PIN)
85 		*s++ = 'p';
86 
87 	c = (caps >> CEPH_CAP_SAUTH) & 3;
88 	if (c) {
89 		*s++ = 'A';
90 		s = gcap_string(s, c);
91 	}
92 
93 	c = (caps >> CEPH_CAP_SLINK) & 3;
94 	if (c) {
95 		*s++ = 'L';
96 		s = gcap_string(s, c);
97 	}
98 
99 	c = (caps >> CEPH_CAP_SXATTR) & 3;
100 	if (c) {
101 		*s++ = 'X';
102 		s = gcap_string(s, c);
103 	}
104 
105 	c = caps >> CEPH_CAP_SFILE;
106 	if (c) {
107 		*s++ = 'F';
108 		s = gcap_string(s, c);
109 	}
110 
111 	if (s == cap_str[i])
112 		*s++ = '-';
113 	*s = 0;
114 	return cap_str[i];
115 }
116 
ceph_caps_init(struct ceph_mds_client * mdsc)117 void ceph_caps_init(struct ceph_mds_client *mdsc)
118 {
119 	INIT_LIST_HEAD(&mdsc->caps_list);
120 	spin_lock_init(&mdsc->caps_list_lock);
121 }
122 
ceph_caps_finalize(struct ceph_mds_client * mdsc)123 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
124 {
125 	struct ceph_cap *cap;
126 
127 	spin_lock(&mdsc->caps_list_lock);
128 	while (!list_empty(&mdsc->caps_list)) {
129 		cap = list_first_entry(&mdsc->caps_list,
130 				       struct ceph_cap, caps_item);
131 		list_del(&cap->caps_item);
132 		kmem_cache_free(ceph_cap_cachep, cap);
133 	}
134 	mdsc->caps_total_count = 0;
135 	mdsc->caps_avail_count = 0;
136 	mdsc->caps_use_count = 0;
137 	mdsc->caps_reserve_count = 0;
138 	mdsc->caps_min_count = 0;
139 	spin_unlock(&mdsc->caps_list_lock);
140 }
141 
ceph_adjust_min_caps(struct ceph_mds_client * mdsc,int delta)142 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
143 {
144 	spin_lock(&mdsc->caps_list_lock);
145 	mdsc->caps_min_count += delta;
146 	BUG_ON(mdsc->caps_min_count < 0);
147 	spin_unlock(&mdsc->caps_list_lock);
148 }
149 
ceph_reserve_caps(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx,int need)150 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
151 		      struct ceph_cap_reservation *ctx, int need)
152 {
153 	int i;
154 	struct ceph_cap *cap;
155 	int have;
156 	int alloc = 0;
157 	LIST_HEAD(newcaps);
158 	int ret = 0;
159 
160 	dout("reserve caps ctx=%p need=%d\n", ctx, need);
161 
162 	/* first reserve any caps that are already allocated */
163 	spin_lock(&mdsc->caps_list_lock);
164 	if (mdsc->caps_avail_count >= need)
165 		have = need;
166 	else
167 		have = mdsc->caps_avail_count;
168 	mdsc->caps_avail_count -= have;
169 	mdsc->caps_reserve_count += have;
170 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 					 mdsc->caps_reserve_count +
172 					 mdsc->caps_avail_count);
173 	spin_unlock(&mdsc->caps_list_lock);
174 
175 	for (i = have; i < need; i++) {
176 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177 		if (!cap) {
178 			ret = -ENOMEM;
179 			goto out_alloc_count;
180 		}
181 		list_add(&cap->caps_item, &newcaps);
182 		alloc++;
183 	}
184 	BUG_ON(have + alloc != need);
185 
186 	spin_lock(&mdsc->caps_list_lock);
187 	mdsc->caps_total_count += alloc;
188 	mdsc->caps_reserve_count += alloc;
189 	list_splice(&newcaps, &mdsc->caps_list);
190 
191 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 					 mdsc->caps_reserve_count +
193 					 mdsc->caps_avail_count);
194 	spin_unlock(&mdsc->caps_list_lock);
195 
196 	ctx->count = need;
197 	dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
198 	     ctx, mdsc->caps_total_count, mdsc->caps_use_count,
199 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
200 	return 0;
201 
202 out_alloc_count:
203 	/* we didn't manage to reserve as much as we needed */
204 	pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
205 		   ctx, need, have);
206 	return ret;
207 }
208 
ceph_unreserve_caps(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx)209 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
210 			struct ceph_cap_reservation *ctx)
211 {
212 	dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
213 	if (ctx->count) {
214 		spin_lock(&mdsc->caps_list_lock);
215 		BUG_ON(mdsc->caps_reserve_count < ctx->count);
216 		mdsc->caps_reserve_count -= ctx->count;
217 		mdsc->caps_avail_count += ctx->count;
218 		ctx->count = 0;
219 		dout("unreserve caps %d = %d used + %d resv + %d avail\n",
220 		     mdsc->caps_total_count, mdsc->caps_use_count,
221 		     mdsc->caps_reserve_count, mdsc->caps_avail_count);
222 		BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
223 						 mdsc->caps_reserve_count +
224 						 mdsc->caps_avail_count);
225 		spin_unlock(&mdsc->caps_list_lock);
226 	}
227 	return 0;
228 }
229 
get_cap(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx)230 static struct ceph_cap *get_cap(struct ceph_mds_client *mdsc,
231 				struct ceph_cap_reservation *ctx)
232 {
233 	struct ceph_cap *cap = NULL;
234 
235 	/* temporary, until we do something about cap import/export */
236 	if (!ctx) {
237 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 		if (cap) {
239 			mdsc->caps_use_count++;
240 			mdsc->caps_total_count++;
241 		}
242 		return cap;
243 	}
244 
245 	spin_lock(&mdsc->caps_list_lock);
246 	dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
247 	     ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
248 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
249 	BUG_ON(!ctx->count);
250 	BUG_ON(ctx->count > mdsc->caps_reserve_count);
251 	BUG_ON(list_empty(&mdsc->caps_list));
252 
253 	ctx->count--;
254 	mdsc->caps_reserve_count--;
255 	mdsc->caps_use_count++;
256 
257 	cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
258 	list_del(&cap->caps_item);
259 
260 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
261 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
262 	spin_unlock(&mdsc->caps_list_lock);
263 	return cap;
264 }
265 
ceph_put_cap(struct ceph_mds_client * mdsc,struct ceph_cap * cap)266 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
267 {
268 	spin_lock(&mdsc->caps_list_lock);
269 	dout("put_cap %p %d = %d used + %d resv + %d avail\n",
270 	     cap, mdsc->caps_total_count, mdsc->caps_use_count,
271 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
272 	mdsc->caps_use_count--;
273 	/*
274 	 * Keep some preallocated caps around (ceph_min_count), to
275 	 * avoid lots of free/alloc churn.
276 	 */
277 	if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
278 				      mdsc->caps_min_count) {
279 		mdsc->caps_total_count--;
280 		kmem_cache_free(ceph_cap_cachep, cap);
281 	} else {
282 		mdsc->caps_avail_count++;
283 		list_add(&cap->caps_item, &mdsc->caps_list);
284 	}
285 
286 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
287 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
288 	spin_unlock(&mdsc->caps_list_lock);
289 }
290 
ceph_reservation_status(struct ceph_fs_client * fsc,int * total,int * avail,int * used,int * reserved,int * min)291 void ceph_reservation_status(struct ceph_fs_client *fsc,
292 			     int *total, int *avail, int *used, int *reserved,
293 			     int *min)
294 {
295 	struct ceph_mds_client *mdsc = fsc->mdsc;
296 
297 	if (total)
298 		*total = mdsc->caps_total_count;
299 	if (avail)
300 		*avail = mdsc->caps_avail_count;
301 	if (used)
302 		*used = mdsc->caps_use_count;
303 	if (reserved)
304 		*reserved = mdsc->caps_reserve_count;
305 	if (min)
306 		*min = mdsc->caps_min_count;
307 }
308 
309 /*
310  * Find ceph_cap for given mds, if any.
311  *
312  * Called with i_ceph_lock held.
313  */
__get_cap_for_mds(struct ceph_inode_info * ci,int mds)314 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
315 {
316 	struct ceph_cap *cap;
317 	struct rb_node *n = ci->i_caps.rb_node;
318 
319 	while (n) {
320 		cap = rb_entry(n, struct ceph_cap, ci_node);
321 		if (mds < cap->mds)
322 			n = n->rb_left;
323 		else if (mds > cap->mds)
324 			n = n->rb_right;
325 		else
326 			return cap;
327 	}
328 	return NULL;
329 }
330 
ceph_get_cap_for_mds(struct ceph_inode_info * ci,int mds)331 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
332 {
333 	struct ceph_cap *cap;
334 
335 	spin_lock(&ci->i_ceph_lock);
336 	cap = __get_cap_for_mds(ci, mds);
337 	spin_unlock(&ci->i_ceph_lock);
338 	return cap;
339 }
340 
341 /*
342  * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
343  */
__ceph_get_cap_mds(struct ceph_inode_info * ci)344 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
345 {
346 	struct ceph_cap *cap;
347 	int mds = -1;
348 	struct rb_node *p;
349 
350 	/* prefer mds with WR|BUFFER|EXCL caps */
351 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
352 		cap = rb_entry(p, struct ceph_cap, ci_node);
353 		mds = cap->mds;
354 		if (cap->issued & (CEPH_CAP_FILE_WR |
355 				   CEPH_CAP_FILE_BUFFER |
356 				   CEPH_CAP_FILE_EXCL))
357 			break;
358 	}
359 	return mds;
360 }
361 
ceph_get_cap_mds(struct inode * inode)362 int ceph_get_cap_mds(struct inode *inode)
363 {
364 	struct ceph_inode_info *ci = ceph_inode(inode);
365 	int mds;
366 	spin_lock(&ci->i_ceph_lock);
367 	mds = __ceph_get_cap_mds(ceph_inode(inode));
368 	spin_unlock(&ci->i_ceph_lock);
369 	return mds;
370 }
371 
372 /*
373  * Called under i_ceph_lock.
374  */
__insert_cap_node(struct ceph_inode_info * ci,struct ceph_cap * new)375 static void __insert_cap_node(struct ceph_inode_info *ci,
376 			      struct ceph_cap *new)
377 {
378 	struct rb_node **p = &ci->i_caps.rb_node;
379 	struct rb_node *parent = NULL;
380 	struct ceph_cap *cap = NULL;
381 
382 	while (*p) {
383 		parent = *p;
384 		cap = rb_entry(parent, struct ceph_cap, ci_node);
385 		if (new->mds < cap->mds)
386 			p = &(*p)->rb_left;
387 		else if (new->mds > cap->mds)
388 			p = &(*p)->rb_right;
389 		else
390 			BUG();
391 	}
392 
393 	rb_link_node(&new->ci_node, parent, p);
394 	rb_insert_color(&new->ci_node, &ci->i_caps);
395 }
396 
397 /*
398  * (re)set cap hold timeouts, which control the delayed release
399  * of unused caps back to the MDS.  Should be called on cap use.
400  */
__cap_set_timeouts(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)401 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
402 			       struct ceph_inode_info *ci)
403 {
404 	struct ceph_mount_options *ma = mdsc->fsc->mount_options;
405 
406 	ci->i_hold_caps_min = round_jiffies(jiffies +
407 					    ma->caps_wanted_delay_min * HZ);
408 	ci->i_hold_caps_max = round_jiffies(jiffies +
409 					    ma->caps_wanted_delay_max * HZ);
410 	dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
411 	     ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
412 }
413 
414 /*
415  * (Re)queue cap at the end of the delayed cap release list.
416  *
417  * If I_FLUSH is set, leave the inode at the front of the list.
418  *
419  * Caller holds i_ceph_lock
420  *    -> we take mdsc->cap_delay_lock
421  */
__cap_delay_requeue(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)422 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
423 				struct ceph_inode_info *ci)
424 {
425 	__cap_set_timeouts(mdsc, ci);
426 	dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
427 	     ci->i_ceph_flags, ci->i_hold_caps_max);
428 	if (!mdsc->stopping) {
429 		spin_lock(&mdsc->cap_delay_lock);
430 		if (!list_empty(&ci->i_cap_delay_list)) {
431 			if (ci->i_ceph_flags & CEPH_I_FLUSH)
432 				goto no_change;
433 			list_del_init(&ci->i_cap_delay_list);
434 		}
435 		list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
436 no_change:
437 		spin_unlock(&mdsc->cap_delay_lock);
438 	}
439 }
440 
441 /*
442  * Queue an inode for immediate writeback.  Mark inode with I_FLUSH,
443  * indicating we should send a cap message to flush dirty metadata
444  * asap, and move to the front of the delayed cap list.
445  */
__cap_delay_requeue_front(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)446 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
447 				      struct ceph_inode_info *ci)
448 {
449 	dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
450 	spin_lock(&mdsc->cap_delay_lock);
451 	ci->i_ceph_flags |= CEPH_I_FLUSH;
452 	if (!list_empty(&ci->i_cap_delay_list))
453 		list_del_init(&ci->i_cap_delay_list);
454 	list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
455 	spin_unlock(&mdsc->cap_delay_lock);
456 }
457 
458 /*
459  * Cancel delayed work on cap.
460  *
461  * Caller must hold i_ceph_lock.
462  */
__cap_delay_cancel(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)463 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
464 			       struct ceph_inode_info *ci)
465 {
466 	dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
467 	if (list_empty(&ci->i_cap_delay_list))
468 		return;
469 	spin_lock(&mdsc->cap_delay_lock);
470 	list_del_init(&ci->i_cap_delay_list);
471 	spin_unlock(&mdsc->cap_delay_lock);
472 }
473 
474 /*
475  * Common issue checks for add_cap, handle_cap_grant.
476  */
__check_cap_issue(struct ceph_inode_info * ci,struct ceph_cap * cap,unsigned issued)477 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
478 			      unsigned issued)
479 {
480 	unsigned had = __ceph_caps_issued(ci, NULL);
481 
482 	/*
483 	 * Each time we receive FILE_CACHE anew, we increment
484 	 * i_rdcache_gen.
485 	 */
486 	if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
487 	    (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
488 		ci->i_rdcache_gen++;
489 
490 	/*
491 	 * if we are newly issued FILE_SHARED, clear D_COMPLETE; we
492 	 * don't know what happened to this directory while we didn't
493 	 * have the cap.
494 	 */
495 	if ((issued & CEPH_CAP_FILE_SHARED) &&
496 	    (had & CEPH_CAP_FILE_SHARED) == 0) {
497 		ci->i_shared_gen++;
498 		if (S_ISDIR(ci->vfs_inode.i_mode))
499 			ceph_dir_clear_complete(&ci->vfs_inode);
500 	}
501 }
502 
503 /*
504  * Add a capability under the given MDS session.
505  *
506  * Caller should hold session snap_rwsem (read) and s_mutex.
507  *
508  * @fmode is the open file mode, if we are opening a file, otherwise
509  * it is < 0.  (This is so we can atomically add the cap and add an
510  * open file reference to it.)
511  */
ceph_add_cap(struct inode * inode,struct ceph_mds_session * session,u64 cap_id,int fmode,unsigned issued,unsigned wanted,unsigned seq,unsigned mseq,u64 realmino,int flags,struct ceph_cap_reservation * caps_reservation)512 int ceph_add_cap(struct inode *inode,
513 		 struct ceph_mds_session *session, u64 cap_id,
514 		 int fmode, unsigned issued, unsigned wanted,
515 		 unsigned seq, unsigned mseq, u64 realmino, int flags,
516 		 struct ceph_cap_reservation *caps_reservation)
517 {
518 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
519 	struct ceph_inode_info *ci = ceph_inode(inode);
520 	struct ceph_cap *new_cap = NULL;
521 	struct ceph_cap *cap;
522 	int mds = session->s_mds;
523 	int actual_wanted;
524 
525 	dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
526 	     session->s_mds, cap_id, ceph_cap_string(issued), seq);
527 
528 	/*
529 	 * If we are opening the file, include file mode wanted bits
530 	 * in wanted.
531 	 */
532 	if (fmode >= 0)
533 		wanted |= ceph_caps_for_mode(fmode);
534 
535 retry:
536 	spin_lock(&ci->i_ceph_lock);
537 	cap = __get_cap_for_mds(ci, mds);
538 	if (!cap) {
539 		if (new_cap) {
540 			cap = new_cap;
541 			new_cap = NULL;
542 		} else {
543 			spin_unlock(&ci->i_ceph_lock);
544 			new_cap = get_cap(mdsc, caps_reservation);
545 			if (new_cap == NULL)
546 				return -ENOMEM;
547 			goto retry;
548 		}
549 
550 		cap->issued = 0;
551 		cap->implemented = 0;
552 		cap->mds = mds;
553 		cap->mds_wanted = 0;
554 
555 		cap->ci = ci;
556 		__insert_cap_node(ci, cap);
557 
558 		/* clear out old exporting info?  (i.e. on cap import) */
559 		if (ci->i_cap_exporting_mds == mds) {
560 			ci->i_cap_exporting_issued = 0;
561 			ci->i_cap_exporting_mseq = 0;
562 			ci->i_cap_exporting_mds = -1;
563 		}
564 
565 		/* add to session cap list */
566 		cap->session = session;
567 		spin_lock(&session->s_cap_lock);
568 		list_add_tail(&cap->session_caps, &session->s_caps);
569 		session->s_nr_caps++;
570 		spin_unlock(&session->s_cap_lock);
571 	} else if (new_cap)
572 		ceph_put_cap(mdsc, new_cap);
573 
574 	if (!ci->i_snap_realm) {
575 		/*
576 		 * add this inode to the appropriate snap realm
577 		 */
578 		struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
579 							       realmino);
580 		if (realm) {
581 			ceph_get_snap_realm(mdsc, realm);
582 			spin_lock(&realm->inodes_with_caps_lock);
583 			ci->i_snap_realm = realm;
584 			list_add(&ci->i_snap_realm_item,
585 				 &realm->inodes_with_caps);
586 			spin_unlock(&realm->inodes_with_caps_lock);
587 		} else {
588 			pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
589 			       realmino);
590 			WARN_ON(!realm);
591 		}
592 	}
593 
594 	__check_cap_issue(ci, cap, issued);
595 
596 	/*
597 	 * If we are issued caps we don't want, or the mds' wanted
598 	 * value appears to be off, queue a check so we'll release
599 	 * later and/or update the mds wanted value.
600 	 */
601 	actual_wanted = __ceph_caps_wanted(ci);
602 	if ((wanted & ~actual_wanted) ||
603 	    (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
604 		dout(" issued %s, mds wanted %s, actual %s, queueing\n",
605 		     ceph_cap_string(issued), ceph_cap_string(wanted),
606 		     ceph_cap_string(actual_wanted));
607 		__cap_delay_requeue(mdsc, ci);
608 	}
609 
610 	if (flags & CEPH_CAP_FLAG_AUTH)
611 		ci->i_auth_cap = cap;
612 	else if (ci->i_auth_cap == cap)
613 		ci->i_auth_cap = NULL;
614 
615 	dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
616 	     inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
617 	     ceph_cap_string(issued|cap->issued), seq, mds);
618 	cap->cap_id = cap_id;
619 	cap->issued = issued;
620 	cap->implemented |= issued;
621 	cap->mds_wanted |= wanted;
622 	cap->seq = seq;
623 	cap->issue_seq = seq;
624 	cap->mseq = mseq;
625 	cap->cap_gen = session->s_cap_gen;
626 
627 	if (fmode >= 0)
628 		__ceph_get_fmode(ci, fmode);
629 	spin_unlock(&ci->i_ceph_lock);
630 	wake_up_all(&ci->i_cap_wq);
631 	return 0;
632 }
633 
634 /*
635  * Return true if cap has not timed out and belongs to the current
636  * generation of the MDS session (i.e. has not gone 'stale' due to
637  * us losing touch with the mds).
638  */
__cap_is_valid(struct ceph_cap * cap)639 static int __cap_is_valid(struct ceph_cap *cap)
640 {
641 	unsigned long ttl;
642 	u32 gen;
643 
644 	spin_lock(&cap->session->s_gen_ttl_lock);
645 	gen = cap->session->s_cap_gen;
646 	ttl = cap->session->s_cap_ttl;
647 	spin_unlock(&cap->session->s_gen_ttl_lock);
648 
649 	if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
650 		dout("__cap_is_valid %p cap %p issued %s "
651 		     "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
652 		     cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
653 		return 0;
654 	}
655 
656 	return 1;
657 }
658 
659 /*
660  * Return set of valid cap bits issued to us.  Note that caps time
661  * out, and may be invalidated in bulk if the client session times out
662  * and session->s_cap_gen is bumped.
663  */
__ceph_caps_issued(struct ceph_inode_info * ci,int * implemented)664 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
665 {
666 	int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
667 	struct ceph_cap *cap;
668 	struct rb_node *p;
669 
670 	if (implemented)
671 		*implemented = 0;
672 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
673 		cap = rb_entry(p, struct ceph_cap, ci_node);
674 		if (!__cap_is_valid(cap))
675 			continue;
676 		dout("__ceph_caps_issued %p cap %p issued %s\n",
677 		     &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
678 		have |= cap->issued;
679 		if (implemented)
680 			*implemented |= cap->implemented;
681 	}
682 	return have;
683 }
684 
685 /*
686  * Get cap bits issued by caps other than @ocap
687  */
__ceph_caps_issued_other(struct ceph_inode_info * ci,struct ceph_cap * ocap)688 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
689 {
690 	int have = ci->i_snap_caps;
691 	struct ceph_cap *cap;
692 	struct rb_node *p;
693 
694 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
695 		cap = rb_entry(p, struct ceph_cap, ci_node);
696 		if (cap == ocap)
697 			continue;
698 		if (!__cap_is_valid(cap))
699 			continue;
700 		have |= cap->issued;
701 	}
702 	return have;
703 }
704 
705 /*
706  * Move a cap to the end of the LRU (oldest caps at list head, newest
707  * at list tail).
708  */
__touch_cap(struct ceph_cap * cap)709 static void __touch_cap(struct ceph_cap *cap)
710 {
711 	struct ceph_mds_session *s = cap->session;
712 
713 	spin_lock(&s->s_cap_lock);
714 	if (s->s_cap_iterator == NULL) {
715 		dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
716 		     s->s_mds);
717 		list_move_tail(&cap->session_caps, &s->s_caps);
718 	} else {
719 		dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
720 		     &cap->ci->vfs_inode, cap, s->s_mds);
721 	}
722 	spin_unlock(&s->s_cap_lock);
723 }
724 
725 /*
726  * Check if we hold the given mask.  If so, move the cap(s) to the
727  * front of their respective LRUs.  (This is the preferred way for
728  * callers to check for caps they want.)
729  */
__ceph_caps_issued_mask(struct ceph_inode_info * ci,int mask,int touch)730 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
731 {
732 	struct ceph_cap *cap;
733 	struct rb_node *p;
734 	int have = ci->i_snap_caps;
735 
736 	if ((have & mask) == mask) {
737 		dout("__ceph_caps_issued_mask %p snap issued %s"
738 		     " (mask %s)\n", &ci->vfs_inode,
739 		     ceph_cap_string(have),
740 		     ceph_cap_string(mask));
741 		return 1;
742 	}
743 
744 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
745 		cap = rb_entry(p, struct ceph_cap, ci_node);
746 		if (!__cap_is_valid(cap))
747 			continue;
748 		if ((cap->issued & mask) == mask) {
749 			dout("__ceph_caps_issued_mask %p cap %p issued %s"
750 			     " (mask %s)\n", &ci->vfs_inode, cap,
751 			     ceph_cap_string(cap->issued),
752 			     ceph_cap_string(mask));
753 			if (touch)
754 				__touch_cap(cap);
755 			return 1;
756 		}
757 
758 		/* does a combination of caps satisfy mask? */
759 		have |= cap->issued;
760 		if ((have & mask) == mask) {
761 			dout("__ceph_caps_issued_mask %p combo issued %s"
762 			     " (mask %s)\n", &ci->vfs_inode,
763 			     ceph_cap_string(cap->issued),
764 			     ceph_cap_string(mask));
765 			if (touch) {
766 				struct rb_node *q;
767 
768 				/* touch this + preceding caps */
769 				__touch_cap(cap);
770 				for (q = rb_first(&ci->i_caps); q != p;
771 				     q = rb_next(q)) {
772 					cap = rb_entry(q, struct ceph_cap,
773 						       ci_node);
774 					if (!__cap_is_valid(cap))
775 						continue;
776 					__touch_cap(cap);
777 				}
778 			}
779 			return 1;
780 		}
781 	}
782 
783 	return 0;
784 }
785 
786 /*
787  * Return true if mask caps are currently being revoked by an MDS.
788  */
ceph_caps_revoking(struct ceph_inode_info * ci,int mask)789 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
790 {
791 	struct inode *inode = &ci->vfs_inode;
792 	struct ceph_cap *cap;
793 	struct rb_node *p;
794 	int ret = 0;
795 
796 	spin_lock(&ci->i_ceph_lock);
797 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
798 		cap = rb_entry(p, struct ceph_cap, ci_node);
799 		if (__cap_is_valid(cap) &&
800 		    (cap->implemented & ~cap->issued & mask)) {
801 			ret = 1;
802 			break;
803 		}
804 	}
805 	spin_unlock(&ci->i_ceph_lock);
806 	dout("ceph_caps_revoking %p %s = %d\n", inode,
807 	     ceph_cap_string(mask), ret);
808 	return ret;
809 }
810 
__ceph_caps_used(struct ceph_inode_info * ci)811 int __ceph_caps_used(struct ceph_inode_info *ci)
812 {
813 	int used = 0;
814 	if (ci->i_pin_ref)
815 		used |= CEPH_CAP_PIN;
816 	if (ci->i_rd_ref)
817 		used |= CEPH_CAP_FILE_RD;
818 	if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
819 		used |= CEPH_CAP_FILE_CACHE;
820 	if (ci->i_wr_ref)
821 		used |= CEPH_CAP_FILE_WR;
822 	if (ci->i_wb_ref || ci->i_wrbuffer_ref)
823 		used |= CEPH_CAP_FILE_BUFFER;
824 	return used;
825 }
826 
827 /*
828  * wanted, by virtue of open file modes
829  */
__ceph_caps_file_wanted(struct ceph_inode_info * ci)830 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
831 {
832 	int want = 0;
833 	int mode;
834 	for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
835 		if (ci->i_nr_by_mode[mode])
836 			want |= ceph_caps_for_mode(mode);
837 	return want;
838 }
839 
840 /*
841  * Return caps we have registered with the MDS(s) as 'wanted'.
842  */
__ceph_caps_mds_wanted(struct ceph_inode_info * ci)843 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
844 {
845 	struct ceph_cap *cap;
846 	struct rb_node *p;
847 	int mds_wanted = 0;
848 
849 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
850 		cap = rb_entry(p, struct ceph_cap, ci_node);
851 		if (!__cap_is_valid(cap))
852 			continue;
853 		mds_wanted |= cap->mds_wanted;
854 	}
855 	return mds_wanted;
856 }
857 
858 /*
859  * called under i_ceph_lock
860  */
__ceph_is_any_caps(struct ceph_inode_info * ci)861 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
862 {
863 	return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
864 }
865 
866 /*
867  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
868  *
869  * caller should hold i_ceph_lock.
870  * caller will not hold session s_mutex if called from destroy_inode.
871  */
__ceph_remove_cap(struct ceph_cap * cap)872 void __ceph_remove_cap(struct ceph_cap *cap)
873 {
874 	struct ceph_mds_session *session = cap->session;
875 	struct ceph_inode_info *ci = cap->ci;
876 	struct ceph_mds_client *mdsc =
877 		ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
878 	int removed = 0;
879 
880 	dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
881 
882 	/* remove from session list */
883 	spin_lock(&session->s_cap_lock);
884 	if (session->s_cap_iterator == cap) {
885 		/* not yet, we are iterating over this very cap */
886 		dout("__ceph_remove_cap  delaying %p removal from session %p\n",
887 		     cap, cap->session);
888 	} else {
889 		list_del_init(&cap->session_caps);
890 		session->s_nr_caps--;
891 		cap->session = NULL;
892 		removed = 1;
893 	}
894 	/* protect backpointer with s_cap_lock: see iterate_session_caps */
895 	cap->ci = NULL;
896 	spin_unlock(&session->s_cap_lock);
897 
898 	/* remove from inode list */
899 	rb_erase(&cap->ci_node, &ci->i_caps);
900 	if (ci->i_auth_cap == cap)
901 		ci->i_auth_cap = NULL;
902 
903 	if (removed)
904 		ceph_put_cap(mdsc, cap);
905 
906 	if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
907 		struct ceph_snap_realm *realm = ci->i_snap_realm;
908 		spin_lock(&realm->inodes_with_caps_lock);
909 		list_del_init(&ci->i_snap_realm_item);
910 		ci->i_snap_realm_counter++;
911 		ci->i_snap_realm = NULL;
912 		spin_unlock(&realm->inodes_with_caps_lock);
913 		ceph_put_snap_realm(mdsc, realm);
914 	}
915 	if (!__ceph_is_any_real_caps(ci))
916 		__cap_delay_cancel(mdsc, ci);
917 }
918 
919 /*
920  * Build and send a cap message to the given MDS.
921  *
922  * Caller should be holding s_mutex.
923  */
send_cap_msg(struct ceph_mds_session * session,u64 ino,u64 cid,int op,int caps,int wanted,int dirty,u32 seq,u64 flush_tid,u32 issue_seq,u32 mseq,u64 size,u64 max_size,struct timespec * mtime,struct timespec * atime,u64 time_warp_seq,uid_t uid,gid_t gid,umode_t mode,u64 xattr_version,struct ceph_buffer * xattrs_buf,u64 follows)924 static int send_cap_msg(struct ceph_mds_session *session,
925 			u64 ino, u64 cid, int op,
926 			int caps, int wanted, int dirty,
927 			u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
928 			u64 size, u64 max_size,
929 			struct timespec *mtime, struct timespec *atime,
930 			u64 time_warp_seq,
931 			uid_t uid, gid_t gid, umode_t mode,
932 			u64 xattr_version,
933 			struct ceph_buffer *xattrs_buf,
934 			u64 follows)
935 {
936 	struct ceph_mds_caps *fc;
937 	struct ceph_msg *msg;
938 
939 	dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
940 	     " seq %u/%u mseq %u follows %lld size %llu/%llu"
941 	     " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
942 	     cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
943 	     ceph_cap_string(dirty),
944 	     seq, issue_seq, mseq, follows, size, max_size,
945 	     xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
946 
947 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), GFP_NOFS, false);
948 	if (!msg)
949 		return -ENOMEM;
950 
951 	msg->hdr.tid = cpu_to_le64(flush_tid);
952 
953 	fc = msg->front.iov_base;
954 	memset(fc, 0, sizeof(*fc));
955 
956 	fc->cap_id = cpu_to_le64(cid);
957 	fc->op = cpu_to_le32(op);
958 	fc->seq = cpu_to_le32(seq);
959 	fc->issue_seq = cpu_to_le32(issue_seq);
960 	fc->migrate_seq = cpu_to_le32(mseq);
961 	fc->caps = cpu_to_le32(caps);
962 	fc->wanted = cpu_to_le32(wanted);
963 	fc->dirty = cpu_to_le32(dirty);
964 	fc->ino = cpu_to_le64(ino);
965 	fc->snap_follows = cpu_to_le64(follows);
966 
967 	fc->size = cpu_to_le64(size);
968 	fc->max_size = cpu_to_le64(max_size);
969 	if (mtime)
970 		ceph_encode_timespec(&fc->mtime, mtime);
971 	if (atime)
972 		ceph_encode_timespec(&fc->atime, atime);
973 	fc->time_warp_seq = cpu_to_le32(time_warp_seq);
974 
975 	fc->uid = cpu_to_le32(uid);
976 	fc->gid = cpu_to_le32(gid);
977 	fc->mode = cpu_to_le32(mode);
978 
979 	fc->xattr_version = cpu_to_le64(xattr_version);
980 	if (xattrs_buf) {
981 		msg->middle = ceph_buffer_get(xattrs_buf);
982 		fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
983 		msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
984 	}
985 
986 	ceph_con_send(&session->s_con, msg);
987 	return 0;
988 }
989 
__queue_cap_release(struct ceph_mds_session * session,u64 ino,u64 cap_id,u32 migrate_seq,u32 issue_seq)990 static void __queue_cap_release(struct ceph_mds_session *session,
991 				u64 ino, u64 cap_id, u32 migrate_seq,
992 				u32 issue_seq)
993 {
994 	struct ceph_msg *msg;
995 	struct ceph_mds_cap_release *head;
996 	struct ceph_mds_cap_item *item;
997 
998 	spin_lock(&session->s_cap_lock);
999 	BUG_ON(!session->s_num_cap_releases);
1000 	msg = list_first_entry(&session->s_cap_releases,
1001 			       struct ceph_msg, list_head);
1002 
1003 	dout(" adding %llx release to mds%d msg %p (%d left)\n",
1004 	     ino, session->s_mds, msg, session->s_num_cap_releases);
1005 
1006 	BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1007 	head = msg->front.iov_base;
1008 	head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1009 	item = msg->front.iov_base + msg->front.iov_len;
1010 	item->ino = cpu_to_le64(ino);
1011 	item->cap_id = cpu_to_le64(cap_id);
1012 	item->migrate_seq = cpu_to_le32(migrate_seq);
1013 	item->seq = cpu_to_le32(issue_seq);
1014 
1015 	session->s_num_cap_releases--;
1016 
1017 	msg->front.iov_len += sizeof(*item);
1018 	if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1019 		dout(" release msg %p full\n", msg);
1020 		list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1021 	} else {
1022 		dout(" release msg %p at %d/%d (%d)\n", msg,
1023 		     (int)le32_to_cpu(head->num),
1024 		     (int)CEPH_CAPS_PER_RELEASE,
1025 		     (int)msg->front.iov_len);
1026 	}
1027 	spin_unlock(&session->s_cap_lock);
1028 }
1029 
1030 /*
1031  * Queue cap releases when an inode is dropped from our cache.  Since
1032  * inode is about to be destroyed, there is no need for i_ceph_lock.
1033  */
ceph_queue_caps_release(struct inode * inode)1034 void ceph_queue_caps_release(struct inode *inode)
1035 {
1036 	struct ceph_inode_info *ci = ceph_inode(inode);
1037 	struct rb_node *p;
1038 
1039 	p = rb_first(&ci->i_caps);
1040 	while (p) {
1041 		struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1042 		struct ceph_mds_session *session = cap->session;
1043 
1044 		__queue_cap_release(session, ceph_ino(inode), cap->cap_id,
1045 				    cap->mseq, cap->issue_seq);
1046 		p = rb_next(p);
1047 		__ceph_remove_cap(cap);
1048 	}
1049 }
1050 
1051 /*
1052  * Send a cap msg on the given inode.  Update our caps state, then
1053  * drop i_ceph_lock and send the message.
1054  *
1055  * Make note of max_size reported/requested from mds, revoked caps
1056  * that have now been implemented.
1057  *
1058  * Make half-hearted attempt ot to invalidate page cache if we are
1059  * dropping RDCACHE.  Note that this will leave behind locked pages
1060  * that we'll then need to deal with elsewhere.
1061  *
1062  * Return non-zero if delayed release, or we experienced an error
1063  * such that the caller should requeue + retry later.
1064  *
1065  * called with i_ceph_lock, then drops it.
1066  * caller should hold snap_rwsem (read), s_mutex.
1067  */
__send_cap(struct ceph_mds_client * mdsc,struct ceph_cap * cap,int op,int used,int want,int retain,int flushing,unsigned * pflush_tid)1068 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1069 		      int op, int used, int want, int retain, int flushing,
1070 		      unsigned *pflush_tid)
1071 	__releases(cap->ci->i_ceph_lock)
1072 {
1073 	struct ceph_inode_info *ci = cap->ci;
1074 	struct inode *inode = &ci->vfs_inode;
1075 	u64 cap_id = cap->cap_id;
1076 	int held, revoking, dropping, keep;
1077 	u64 seq, issue_seq, mseq, time_warp_seq, follows;
1078 	u64 size, max_size;
1079 	struct timespec mtime, atime;
1080 	int wake = 0;
1081 	umode_t mode;
1082 	uid_t uid;
1083 	gid_t gid;
1084 	struct ceph_mds_session *session;
1085 	u64 xattr_version = 0;
1086 	struct ceph_buffer *xattr_blob = NULL;
1087 	int delayed = 0;
1088 	u64 flush_tid = 0;
1089 	int i;
1090 	int ret;
1091 
1092 	held = cap->issued | cap->implemented;
1093 	revoking = cap->implemented & ~cap->issued;
1094 	retain &= ~revoking;
1095 	dropping = cap->issued & ~retain;
1096 
1097 	dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1098 	     inode, cap, cap->session,
1099 	     ceph_cap_string(held), ceph_cap_string(held & retain),
1100 	     ceph_cap_string(revoking));
1101 	BUG_ON((retain & CEPH_CAP_PIN) == 0);
1102 
1103 	session = cap->session;
1104 
1105 	/* don't release wanted unless we've waited a bit. */
1106 	if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1107 	    time_before(jiffies, ci->i_hold_caps_min)) {
1108 		dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1109 		     ceph_cap_string(cap->issued),
1110 		     ceph_cap_string(cap->issued & retain),
1111 		     ceph_cap_string(cap->mds_wanted),
1112 		     ceph_cap_string(want));
1113 		want |= cap->mds_wanted;
1114 		retain |= cap->issued;
1115 		delayed = 1;
1116 	}
1117 	ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1118 
1119 	cap->issued &= retain;  /* drop bits we don't want */
1120 	if (cap->implemented & ~cap->issued) {
1121 		/*
1122 		 * Wake up any waiters on wanted -> needed transition.
1123 		 * This is due to the weird transition from buffered
1124 		 * to sync IO... we need to flush dirty pages _before_
1125 		 * allowing sync writes to avoid reordering.
1126 		 */
1127 		wake = 1;
1128 	}
1129 	cap->implemented &= cap->issued | used;
1130 	cap->mds_wanted = want;
1131 
1132 	if (flushing) {
1133 		/*
1134 		 * assign a tid for flush operations so we can avoid
1135 		 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1136 		 * clean type races.  track latest tid for every bit
1137 		 * so we can handle flush AxFw, flush Fw, and have the
1138 		 * first ack clean Ax.
1139 		 */
1140 		flush_tid = ++ci->i_cap_flush_last_tid;
1141 		if (pflush_tid)
1142 			*pflush_tid = flush_tid;
1143 		dout(" cap_flush_tid %d\n", (int)flush_tid);
1144 		for (i = 0; i < CEPH_CAP_BITS; i++)
1145 			if (flushing & (1 << i))
1146 				ci->i_cap_flush_tid[i] = flush_tid;
1147 
1148 		follows = ci->i_head_snapc->seq;
1149 	} else {
1150 		follows = 0;
1151 	}
1152 
1153 	keep = cap->implemented;
1154 	seq = cap->seq;
1155 	issue_seq = cap->issue_seq;
1156 	mseq = cap->mseq;
1157 	size = inode->i_size;
1158 	ci->i_reported_size = size;
1159 	max_size = ci->i_wanted_max_size;
1160 	ci->i_requested_max_size = max_size;
1161 	mtime = inode->i_mtime;
1162 	atime = inode->i_atime;
1163 	time_warp_seq = ci->i_time_warp_seq;
1164 	uid = inode->i_uid;
1165 	gid = inode->i_gid;
1166 	mode = inode->i_mode;
1167 
1168 	if (flushing & CEPH_CAP_XATTR_EXCL) {
1169 		__ceph_build_xattrs_blob(ci);
1170 		xattr_blob = ci->i_xattrs.blob;
1171 		xattr_version = ci->i_xattrs.version;
1172 	}
1173 
1174 	spin_unlock(&ci->i_ceph_lock);
1175 
1176 	ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1177 		op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1178 		size, max_size, &mtime, &atime, time_warp_seq,
1179 		uid, gid, mode, xattr_version, xattr_blob,
1180 		follows);
1181 	if (ret < 0) {
1182 		dout("error sending cap msg, must requeue %p\n", inode);
1183 		delayed = 1;
1184 	}
1185 
1186 	if (wake)
1187 		wake_up_all(&ci->i_cap_wq);
1188 
1189 	return delayed;
1190 }
1191 
1192 /*
1193  * When a snapshot is taken, clients accumulate dirty metadata on
1194  * inodes with capabilities in ceph_cap_snaps to describe the file
1195  * state at the time the snapshot was taken.  This must be flushed
1196  * asynchronously back to the MDS once sync writes complete and dirty
1197  * data is written out.
1198  *
1199  * Unless @again is true, skip cap_snaps that were already sent to
1200  * the MDS (i.e., during this session).
1201  *
1202  * Called under i_ceph_lock.  Takes s_mutex as needed.
1203  */
__ceph_flush_snaps(struct ceph_inode_info * ci,struct ceph_mds_session ** psession,int again)1204 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1205 			struct ceph_mds_session **psession,
1206 			int again)
1207 		__releases(ci->i_ceph_lock)
1208 		__acquires(ci->i_ceph_lock)
1209 {
1210 	struct inode *inode = &ci->vfs_inode;
1211 	int mds;
1212 	struct ceph_cap_snap *capsnap;
1213 	u32 mseq;
1214 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1215 	struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1216 						    session->s_mutex */
1217 	u64 next_follows = 0;  /* keep track of how far we've gotten through the
1218 			     i_cap_snaps list, and skip these entries next time
1219 			     around to avoid an infinite loop */
1220 
1221 	if (psession)
1222 		session = *psession;
1223 
1224 	dout("__flush_snaps %p\n", inode);
1225 retry:
1226 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1227 		/* avoid an infiniute loop after retry */
1228 		if (capsnap->follows < next_follows)
1229 			continue;
1230 		/*
1231 		 * we need to wait for sync writes to complete and for dirty
1232 		 * pages to be written out.
1233 		 */
1234 		if (capsnap->dirty_pages || capsnap->writing)
1235 			break;
1236 
1237 		/*
1238 		 * if cap writeback already occurred, we should have dropped
1239 		 * the capsnap in ceph_put_wrbuffer_cap_refs.
1240 		 */
1241 		BUG_ON(capsnap->dirty == 0);
1242 
1243 		/* pick mds, take s_mutex */
1244 		if (ci->i_auth_cap == NULL) {
1245 			dout("no auth cap (migrating?), doing nothing\n");
1246 			goto out;
1247 		}
1248 
1249 		/* only flush each capsnap once */
1250 		if (!again && !list_empty(&capsnap->flushing_item)) {
1251 			dout("already flushed %p, skipping\n", capsnap);
1252 			continue;
1253 		}
1254 
1255 		mds = ci->i_auth_cap->session->s_mds;
1256 		mseq = ci->i_auth_cap->mseq;
1257 
1258 		if (session && session->s_mds != mds) {
1259 			dout("oops, wrong session %p mutex\n", session);
1260 			mutex_unlock(&session->s_mutex);
1261 			ceph_put_mds_session(session);
1262 			session = NULL;
1263 		}
1264 		if (!session) {
1265 			spin_unlock(&ci->i_ceph_lock);
1266 			mutex_lock(&mdsc->mutex);
1267 			session = __ceph_lookup_mds_session(mdsc, mds);
1268 			mutex_unlock(&mdsc->mutex);
1269 			if (session) {
1270 				dout("inverting session/ino locks on %p\n",
1271 				     session);
1272 				mutex_lock(&session->s_mutex);
1273 			}
1274 			/*
1275 			 * if session == NULL, we raced against a cap
1276 			 * deletion or migration.  retry, and we'll
1277 			 * get a better @mds value next time.
1278 			 */
1279 			spin_lock(&ci->i_ceph_lock);
1280 			goto retry;
1281 		}
1282 
1283 		capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1284 		atomic_inc(&capsnap->nref);
1285 		if (!list_empty(&capsnap->flushing_item))
1286 			list_del_init(&capsnap->flushing_item);
1287 		list_add_tail(&capsnap->flushing_item,
1288 			      &session->s_cap_snaps_flushing);
1289 		spin_unlock(&ci->i_ceph_lock);
1290 
1291 		dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1292 		     inode, capsnap, capsnap->follows, capsnap->flush_tid);
1293 		send_cap_msg(session, ceph_vino(inode).ino, 0,
1294 			     CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1295 			     capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1296 			     capsnap->size, 0,
1297 			     &capsnap->mtime, &capsnap->atime,
1298 			     capsnap->time_warp_seq,
1299 			     capsnap->uid, capsnap->gid, capsnap->mode,
1300 			     capsnap->xattr_version, capsnap->xattr_blob,
1301 			     capsnap->follows);
1302 
1303 		next_follows = capsnap->follows + 1;
1304 		ceph_put_cap_snap(capsnap);
1305 
1306 		spin_lock(&ci->i_ceph_lock);
1307 		goto retry;
1308 	}
1309 
1310 	/* we flushed them all; remove this inode from the queue */
1311 	spin_lock(&mdsc->snap_flush_lock);
1312 	list_del_init(&ci->i_snap_flush_item);
1313 	spin_unlock(&mdsc->snap_flush_lock);
1314 
1315 out:
1316 	if (psession)
1317 		*psession = session;
1318 	else if (session) {
1319 		mutex_unlock(&session->s_mutex);
1320 		ceph_put_mds_session(session);
1321 	}
1322 }
1323 
ceph_flush_snaps(struct ceph_inode_info * ci)1324 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1325 {
1326 	spin_lock(&ci->i_ceph_lock);
1327 	__ceph_flush_snaps(ci, NULL, 0);
1328 	spin_unlock(&ci->i_ceph_lock);
1329 }
1330 
1331 /*
1332  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1333  * Caller is then responsible for calling __mark_inode_dirty with the
1334  * returned flags value.
1335  */
__ceph_mark_dirty_caps(struct ceph_inode_info * ci,int mask)1336 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1337 {
1338 	struct ceph_mds_client *mdsc =
1339 		ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1340 	struct inode *inode = &ci->vfs_inode;
1341 	int was = ci->i_dirty_caps;
1342 	int dirty = 0;
1343 
1344 	dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1345 	     ceph_cap_string(mask), ceph_cap_string(was),
1346 	     ceph_cap_string(was | mask));
1347 	ci->i_dirty_caps |= mask;
1348 	if (was == 0) {
1349 		if (!ci->i_head_snapc)
1350 			ci->i_head_snapc = ceph_get_snap_context(
1351 				ci->i_snap_realm->cached_context);
1352 		dout(" inode %p now dirty snapc %p auth cap %p\n",
1353 		     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1354 		BUG_ON(!list_empty(&ci->i_dirty_item));
1355 		spin_lock(&mdsc->cap_dirty_lock);
1356 		if (ci->i_auth_cap)
1357 			list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1358 		else
1359 			list_add(&ci->i_dirty_item,
1360 				 &mdsc->cap_dirty_migrating);
1361 		spin_unlock(&mdsc->cap_dirty_lock);
1362 		if (ci->i_flushing_caps == 0) {
1363 			ihold(inode);
1364 			dirty |= I_DIRTY_SYNC;
1365 		}
1366 	}
1367 	BUG_ON(list_empty(&ci->i_dirty_item));
1368 	if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1369 	    (mask & CEPH_CAP_FILE_BUFFER))
1370 		dirty |= I_DIRTY_DATASYNC;
1371 	__cap_delay_requeue(mdsc, ci);
1372 	return dirty;
1373 }
1374 
1375 /*
1376  * Add dirty inode to the flushing list.  Assigned a seq number so we
1377  * can wait for caps to flush without starving.
1378  *
1379  * Called under i_ceph_lock.
1380  */
__mark_caps_flushing(struct inode * inode,struct ceph_mds_session * session)1381 static int __mark_caps_flushing(struct inode *inode,
1382 				 struct ceph_mds_session *session)
1383 {
1384 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1385 	struct ceph_inode_info *ci = ceph_inode(inode);
1386 	int flushing;
1387 
1388 	BUG_ON(ci->i_dirty_caps == 0);
1389 	BUG_ON(list_empty(&ci->i_dirty_item));
1390 
1391 	flushing = ci->i_dirty_caps;
1392 	dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1393 	     ceph_cap_string(flushing),
1394 	     ceph_cap_string(ci->i_flushing_caps),
1395 	     ceph_cap_string(ci->i_flushing_caps | flushing));
1396 	ci->i_flushing_caps |= flushing;
1397 	ci->i_dirty_caps = 0;
1398 	dout(" inode %p now !dirty\n", inode);
1399 
1400 	spin_lock(&mdsc->cap_dirty_lock);
1401 	list_del_init(&ci->i_dirty_item);
1402 
1403 	ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1404 	if (list_empty(&ci->i_flushing_item)) {
1405 		list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1406 		mdsc->num_cap_flushing++;
1407 		dout(" inode %p now flushing seq %lld\n", inode,
1408 		     ci->i_cap_flush_seq);
1409 	} else {
1410 		list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1411 		dout(" inode %p now flushing (more) seq %lld\n", inode,
1412 		     ci->i_cap_flush_seq);
1413 	}
1414 	spin_unlock(&mdsc->cap_dirty_lock);
1415 
1416 	return flushing;
1417 }
1418 
1419 /*
1420  * try to invalidate mapping pages without blocking.
1421  */
try_nonblocking_invalidate(struct inode * inode)1422 static int try_nonblocking_invalidate(struct inode *inode)
1423 {
1424 	struct ceph_inode_info *ci = ceph_inode(inode);
1425 	u32 invalidating_gen = ci->i_rdcache_gen;
1426 
1427 	spin_unlock(&ci->i_ceph_lock);
1428 	invalidate_mapping_pages(&inode->i_data, 0, -1);
1429 	spin_lock(&ci->i_ceph_lock);
1430 
1431 	if (inode->i_data.nrpages == 0 &&
1432 	    invalidating_gen == ci->i_rdcache_gen) {
1433 		/* success. */
1434 		dout("try_nonblocking_invalidate %p success\n", inode);
1435 		/* save any racing async invalidate some trouble */
1436 		ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1437 		return 0;
1438 	}
1439 	dout("try_nonblocking_invalidate %p failed\n", inode);
1440 	return -1;
1441 }
1442 
1443 /*
1444  * Swiss army knife function to examine currently used and wanted
1445  * versus held caps.  Release, flush, ack revoked caps to mds as
1446  * appropriate.
1447  *
1448  *  CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1449  *    cap release further.
1450  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1451  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1452  *    further delay.
1453  */
ceph_check_caps(struct ceph_inode_info * ci,int flags,struct ceph_mds_session * session)1454 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1455 		     struct ceph_mds_session *session)
1456 {
1457 	struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1458 	struct ceph_mds_client *mdsc = fsc->mdsc;
1459 	struct inode *inode = &ci->vfs_inode;
1460 	struct ceph_cap *cap;
1461 	int file_wanted, used;
1462 	int took_snap_rwsem = 0;             /* true if mdsc->snap_rwsem held */
1463 	int issued, implemented, want, retain, revoking, flushing = 0;
1464 	int mds = -1;   /* keep track of how far we've gone through i_caps list
1465 			   to avoid an infinite loop on retry */
1466 	struct rb_node *p;
1467 	int tried_invalidate = 0;
1468 	int delayed = 0, sent = 0, force_requeue = 0, num;
1469 	int queue_invalidate = 0;
1470 	int is_delayed = flags & CHECK_CAPS_NODELAY;
1471 
1472 	/* if we are unmounting, flush any unused caps immediately. */
1473 	if (mdsc->stopping)
1474 		is_delayed = 1;
1475 
1476 	spin_lock(&ci->i_ceph_lock);
1477 
1478 	if (ci->i_ceph_flags & CEPH_I_FLUSH)
1479 		flags |= CHECK_CAPS_FLUSH;
1480 
1481 	/* flush snaps first time around only */
1482 	if (!list_empty(&ci->i_cap_snaps))
1483 		__ceph_flush_snaps(ci, &session, 0);
1484 	goto retry_locked;
1485 retry:
1486 	spin_lock(&ci->i_ceph_lock);
1487 retry_locked:
1488 	file_wanted = __ceph_caps_file_wanted(ci);
1489 	used = __ceph_caps_used(ci);
1490 	want = file_wanted | used;
1491 	issued = __ceph_caps_issued(ci, &implemented);
1492 	revoking = implemented & ~issued;
1493 
1494 	retain = want | CEPH_CAP_PIN;
1495 	if (!mdsc->stopping && inode->i_nlink > 0) {
1496 		if (want) {
1497 			retain |= CEPH_CAP_ANY;       /* be greedy */
1498 		} else {
1499 			retain |= CEPH_CAP_ANY_SHARED;
1500 			/*
1501 			 * keep RD only if we didn't have the file open RW,
1502 			 * because then the mds would revoke it anyway to
1503 			 * journal max_size=0.
1504 			 */
1505 			if (ci->i_max_size == 0)
1506 				retain |= CEPH_CAP_ANY_RD;
1507 		}
1508 	}
1509 
1510 	dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1511 	     " issued %s revoking %s retain %s %s%s%s\n", inode,
1512 	     ceph_cap_string(file_wanted),
1513 	     ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1514 	     ceph_cap_string(ci->i_flushing_caps),
1515 	     ceph_cap_string(issued), ceph_cap_string(revoking),
1516 	     ceph_cap_string(retain),
1517 	     (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1518 	     (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1519 	     (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1520 
1521 	/*
1522 	 * If we no longer need to hold onto old our caps, and we may
1523 	 * have cached pages, but don't want them, then try to invalidate.
1524 	 * If we fail, it's because pages are locked.... try again later.
1525 	 */
1526 	if ((!is_delayed || mdsc->stopping) &&
1527 	    ci->i_wrbuffer_ref == 0 &&               /* no dirty pages... */
1528 	    inode->i_data.nrpages &&                 /* have cached pages */
1529 	    (file_wanted == 0 ||                     /* no open files */
1530 	     (revoking & (CEPH_CAP_FILE_CACHE|
1531 			  CEPH_CAP_FILE_LAZYIO))) && /*  or revoking cache */
1532 	    !tried_invalidate) {
1533 		dout("check_caps trying to invalidate on %p\n", inode);
1534 		if (try_nonblocking_invalidate(inode) < 0) {
1535 			if (revoking & (CEPH_CAP_FILE_CACHE|
1536 					CEPH_CAP_FILE_LAZYIO)) {
1537 				dout("check_caps queuing invalidate\n");
1538 				queue_invalidate = 1;
1539 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
1540 			} else {
1541 				dout("check_caps failed to invalidate pages\n");
1542 				/* we failed to invalidate pages.  check these
1543 				   caps again later. */
1544 				force_requeue = 1;
1545 				__cap_set_timeouts(mdsc, ci);
1546 			}
1547 		}
1548 		tried_invalidate = 1;
1549 		goto retry_locked;
1550 	}
1551 
1552 	num = 0;
1553 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1554 		cap = rb_entry(p, struct ceph_cap, ci_node);
1555 		num++;
1556 
1557 		/* avoid looping forever */
1558 		if (mds >= cap->mds ||
1559 		    ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1560 			continue;
1561 
1562 		/* NOTE: no side-effects allowed, until we take s_mutex */
1563 
1564 		revoking = cap->implemented & ~cap->issued;
1565 		dout(" mds%d cap %p issued %s implemented %s revoking %s\n",
1566 		     cap->mds, cap, ceph_cap_string(cap->issued),
1567 		     ceph_cap_string(cap->implemented),
1568 		     ceph_cap_string(revoking));
1569 
1570 		if (cap == ci->i_auth_cap &&
1571 		    (cap->issued & CEPH_CAP_FILE_WR)) {
1572 			/* request larger max_size from MDS? */
1573 			if (ci->i_wanted_max_size > ci->i_max_size &&
1574 			    ci->i_wanted_max_size > ci->i_requested_max_size) {
1575 				dout("requesting new max_size\n");
1576 				goto ack;
1577 			}
1578 
1579 			/* approaching file_max? */
1580 			if ((inode->i_size << 1) >= ci->i_max_size &&
1581 			    (ci->i_reported_size << 1) < ci->i_max_size) {
1582 				dout("i_size approaching max_size\n");
1583 				goto ack;
1584 			}
1585 		}
1586 		/* flush anything dirty? */
1587 		if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1588 		    ci->i_dirty_caps) {
1589 			dout("flushing dirty caps\n");
1590 			goto ack;
1591 		}
1592 
1593 		/* completed revocation? going down and there are no caps? */
1594 		if (revoking && (revoking & used) == 0) {
1595 			dout("completed revocation of %s\n",
1596 			     ceph_cap_string(cap->implemented & ~cap->issued));
1597 			goto ack;
1598 		}
1599 
1600 		/* want more caps from mds? */
1601 		if (want & ~(cap->mds_wanted | cap->issued))
1602 			goto ack;
1603 
1604 		/* things we might delay */
1605 		if ((cap->issued & ~retain) == 0 &&
1606 		    cap->mds_wanted == want)
1607 			continue;     /* nope, all good */
1608 
1609 		if (is_delayed)
1610 			goto ack;
1611 
1612 		/* delay? */
1613 		if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1614 		    time_before(jiffies, ci->i_hold_caps_max)) {
1615 			dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1616 			     ceph_cap_string(cap->issued),
1617 			     ceph_cap_string(cap->issued & retain),
1618 			     ceph_cap_string(cap->mds_wanted),
1619 			     ceph_cap_string(want));
1620 			delayed++;
1621 			continue;
1622 		}
1623 
1624 ack:
1625 		if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1626 			dout(" skipping %p I_NOFLUSH set\n", inode);
1627 			continue;
1628 		}
1629 
1630 		if (session && session != cap->session) {
1631 			dout("oops, wrong session %p mutex\n", session);
1632 			mutex_unlock(&session->s_mutex);
1633 			session = NULL;
1634 		}
1635 		if (!session) {
1636 			session = cap->session;
1637 			if (mutex_trylock(&session->s_mutex) == 0) {
1638 				dout("inverting session/ino locks on %p\n",
1639 				     session);
1640 				spin_unlock(&ci->i_ceph_lock);
1641 				if (took_snap_rwsem) {
1642 					up_read(&mdsc->snap_rwsem);
1643 					took_snap_rwsem = 0;
1644 				}
1645 				mutex_lock(&session->s_mutex);
1646 				goto retry;
1647 			}
1648 		}
1649 		/* take snap_rwsem after session mutex */
1650 		if (!took_snap_rwsem) {
1651 			if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1652 				dout("inverting snap/in locks on %p\n",
1653 				     inode);
1654 				spin_unlock(&ci->i_ceph_lock);
1655 				down_read(&mdsc->snap_rwsem);
1656 				took_snap_rwsem = 1;
1657 				goto retry;
1658 			}
1659 			took_snap_rwsem = 1;
1660 		}
1661 
1662 		if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1663 			flushing = __mark_caps_flushing(inode, session);
1664 		else
1665 			flushing = 0;
1666 
1667 		mds = cap->mds;  /* remember mds, so we don't repeat */
1668 		sent++;
1669 
1670 		/* __send_cap drops i_ceph_lock */
1671 		delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1672 				      retain, flushing, NULL);
1673 		goto retry; /* retake i_ceph_lock and restart our cap scan. */
1674 	}
1675 
1676 	/*
1677 	 * Reschedule delayed caps release if we delayed anything,
1678 	 * otherwise cancel.
1679 	 */
1680 	if (delayed && is_delayed)
1681 		force_requeue = 1;   /* __send_cap delayed release; requeue */
1682 	if (!delayed && !is_delayed)
1683 		__cap_delay_cancel(mdsc, ci);
1684 	else if (!is_delayed || force_requeue)
1685 		__cap_delay_requeue(mdsc, ci);
1686 
1687 	spin_unlock(&ci->i_ceph_lock);
1688 
1689 	if (queue_invalidate)
1690 		ceph_queue_invalidate(inode);
1691 
1692 	if (session)
1693 		mutex_unlock(&session->s_mutex);
1694 	if (took_snap_rwsem)
1695 		up_read(&mdsc->snap_rwsem);
1696 }
1697 
1698 /*
1699  * Try to flush dirty caps back to the auth mds.
1700  */
try_flush_caps(struct inode * inode,struct ceph_mds_session * session,unsigned * flush_tid)1701 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1702 			  unsigned *flush_tid)
1703 {
1704 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1705 	struct ceph_inode_info *ci = ceph_inode(inode);
1706 	int unlock_session = session ? 0 : 1;
1707 	int flushing = 0;
1708 
1709 retry:
1710 	spin_lock(&ci->i_ceph_lock);
1711 	if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1712 		dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1713 		goto out;
1714 	}
1715 	if (ci->i_dirty_caps && ci->i_auth_cap) {
1716 		struct ceph_cap *cap = ci->i_auth_cap;
1717 		int used = __ceph_caps_used(ci);
1718 		int want = __ceph_caps_wanted(ci);
1719 		int delayed;
1720 
1721 		if (!session) {
1722 			spin_unlock(&ci->i_ceph_lock);
1723 			session = cap->session;
1724 			mutex_lock(&session->s_mutex);
1725 			goto retry;
1726 		}
1727 		BUG_ON(session != cap->session);
1728 		if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1729 			goto out;
1730 
1731 		flushing = __mark_caps_flushing(inode, session);
1732 
1733 		/* __send_cap drops i_ceph_lock */
1734 		delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1735 				     cap->issued | cap->implemented, flushing,
1736 				     flush_tid);
1737 		if (!delayed)
1738 			goto out_unlocked;
1739 
1740 		spin_lock(&ci->i_ceph_lock);
1741 		__cap_delay_requeue(mdsc, ci);
1742 	}
1743 out:
1744 	spin_unlock(&ci->i_ceph_lock);
1745 out_unlocked:
1746 	if (session && unlock_session)
1747 		mutex_unlock(&session->s_mutex);
1748 	return flushing;
1749 }
1750 
1751 /*
1752  * Return true if we've flushed caps through the given flush_tid.
1753  */
caps_are_flushed(struct inode * inode,unsigned tid)1754 static int caps_are_flushed(struct inode *inode, unsigned tid)
1755 {
1756 	struct ceph_inode_info *ci = ceph_inode(inode);
1757 	int i, ret = 1;
1758 
1759 	spin_lock(&ci->i_ceph_lock);
1760 	for (i = 0; i < CEPH_CAP_BITS; i++)
1761 		if ((ci->i_flushing_caps & (1 << i)) &&
1762 		    ci->i_cap_flush_tid[i] <= tid) {
1763 			/* still flushing this bit */
1764 			ret = 0;
1765 			break;
1766 		}
1767 	spin_unlock(&ci->i_ceph_lock);
1768 	return ret;
1769 }
1770 
1771 /*
1772  * Wait on any unsafe replies for the given inode.  First wait on the
1773  * newest request, and make that the upper bound.  Then, if there are
1774  * more requests, keep waiting on the oldest as long as it is still older
1775  * than the original request.
1776  */
sync_write_wait(struct inode * inode)1777 static void sync_write_wait(struct inode *inode)
1778 {
1779 	struct ceph_inode_info *ci = ceph_inode(inode);
1780 	struct list_head *head = &ci->i_unsafe_writes;
1781 	struct ceph_osd_request *req;
1782 	u64 last_tid;
1783 
1784 	spin_lock(&ci->i_unsafe_lock);
1785 	if (list_empty(head))
1786 		goto out;
1787 
1788 	/* set upper bound as _last_ entry in chain */
1789 	req = list_entry(head->prev, struct ceph_osd_request,
1790 			 r_unsafe_item);
1791 	last_tid = req->r_tid;
1792 
1793 	do {
1794 		ceph_osdc_get_request(req);
1795 		spin_unlock(&ci->i_unsafe_lock);
1796 		dout("sync_write_wait on tid %llu (until %llu)\n",
1797 		     req->r_tid, last_tid);
1798 		wait_for_completion(&req->r_safe_completion);
1799 		spin_lock(&ci->i_unsafe_lock);
1800 		ceph_osdc_put_request(req);
1801 
1802 		/*
1803 		 * from here on look at first entry in chain, since we
1804 		 * only want to wait for anything older than last_tid
1805 		 */
1806 		if (list_empty(head))
1807 			break;
1808 		req = list_entry(head->next, struct ceph_osd_request,
1809 				 r_unsafe_item);
1810 	} while (req->r_tid < last_tid);
1811 out:
1812 	spin_unlock(&ci->i_unsafe_lock);
1813 }
1814 
ceph_fsync(struct file * file,loff_t start,loff_t end,int datasync)1815 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1816 {
1817 	struct inode *inode = file->f_mapping->host;
1818 	struct ceph_inode_info *ci = ceph_inode(inode);
1819 	unsigned flush_tid;
1820 	int ret;
1821 	int dirty;
1822 
1823 	dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1824 	sync_write_wait(inode);
1825 
1826 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1827 	if (ret < 0)
1828 		return ret;
1829 	mutex_lock(&inode->i_mutex);
1830 
1831 	dirty = try_flush_caps(inode, NULL, &flush_tid);
1832 	dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1833 
1834 	/*
1835 	 * only wait on non-file metadata writeback (the mds
1836 	 * can recover size and mtime, so we don't need to
1837 	 * wait for that)
1838 	 */
1839 	if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1840 		dout("fsync waiting for flush_tid %u\n", flush_tid);
1841 		ret = wait_event_interruptible(ci->i_cap_wq,
1842 				       caps_are_flushed(inode, flush_tid));
1843 	}
1844 
1845 	dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1846 	mutex_unlock(&inode->i_mutex);
1847 	return ret;
1848 }
1849 
1850 /*
1851  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
1852  * queue inode for flush but don't do so immediately, because we can
1853  * get by with fewer MDS messages if we wait for data writeback to
1854  * complete first.
1855  */
ceph_write_inode(struct inode * inode,struct writeback_control * wbc)1856 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1857 {
1858 	struct ceph_inode_info *ci = ceph_inode(inode);
1859 	unsigned flush_tid;
1860 	int err = 0;
1861 	int dirty;
1862 	int wait = wbc->sync_mode == WB_SYNC_ALL;
1863 
1864 	dout("write_inode %p wait=%d\n", inode, wait);
1865 	if (wait) {
1866 		dirty = try_flush_caps(inode, NULL, &flush_tid);
1867 		if (dirty)
1868 			err = wait_event_interruptible(ci->i_cap_wq,
1869 				       caps_are_flushed(inode, flush_tid));
1870 	} else {
1871 		struct ceph_mds_client *mdsc =
1872 			ceph_sb_to_client(inode->i_sb)->mdsc;
1873 
1874 		spin_lock(&ci->i_ceph_lock);
1875 		if (__ceph_caps_dirty(ci))
1876 			__cap_delay_requeue_front(mdsc, ci);
1877 		spin_unlock(&ci->i_ceph_lock);
1878 	}
1879 	return err;
1880 }
1881 
1882 /*
1883  * After a recovering MDS goes active, we need to resend any caps
1884  * we were flushing.
1885  *
1886  * Caller holds session->s_mutex.
1887  */
kick_flushing_capsnaps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session)1888 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1889 				   struct ceph_mds_session *session)
1890 {
1891 	struct ceph_cap_snap *capsnap;
1892 
1893 	dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1894 	list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1895 			    flushing_item) {
1896 		struct ceph_inode_info *ci = capsnap->ci;
1897 		struct inode *inode = &ci->vfs_inode;
1898 		struct ceph_cap *cap;
1899 
1900 		spin_lock(&ci->i_ceph_lock);
1901 		cap = ci->i_auth_cap;
1902 		if (cap && cap->session == session) {
1903 			dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1904 			     cap, capsnap);
1905 			__ceph_flush_snaps(ci, &session, 1);
1906 		} else {
1907 			pr_err("%p auth cap %p not mds%d ???\n", inode,
1908 			       cap, session->s_mds);
1909 		}
1910 		spin_unlock(&ci->i_ceph_lock);
1911 	}
1912 }
1913 
ceph_kick_flushing_caps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session)1914 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1915 			     struct ceph_mds_session *session)
1916 {
1917 	struct ceph_inode_info *ci;
1918 
1919 	kick_flushing_capsnaps(mdsc, session);
1920 
1921 	dout("kick_flushing_caps mds%d\n", session->s_mds);
1922 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1923 		struct inode *inode = &ci->vfs_inode;
1924 		struct ceph_cap *cap;
1925 		int delayed = 0;
1926 
1927 		spin_lock(&ci->i_ceph_lock);
1928 		cap = ci->i_auth_cap;
1929 		if (cap && cap->session == session) {
1930 			dout("kick_flushing_caps %p cap %p %s\n", inode,
1931 			     cap, ceph_cap_string(ci->i_flushing_caps));
1932 			delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1933 					     __ceph_caps_used(ci),
1934 					     __ceph_caps_wanted(ci),
1935 					     cap->issued | cap->implemented,
1936 					     ci->i_flushing_caps, NULL);
1937 			if (delayed) {
1938 				spin_lock(&ci->i_ceph_lock);
1939 				__cap_delay_requeue(mdsc, ci);
1940 				spin_unlock(&ci->i_ceph_lock);
1941 			}
1942 		} else {
1943 			pr_err("%p auth cap %p not mds%d ???\n", inode,
1944 			       cap, session->s_mds);
1945 			spin_unlock(&ci->i_ceph_lock);
1946 		}
1947 	}
1948 }
1949 
kick_flushing_inode_caps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session,struct inode * inode)1950 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
1951 				     struct ceph_mds_session *session,
1952 				     struct inode *inode)
1953 {
1954 	struct ceph_inode_info *ci = ceph_inode(inode);
1955 	struct ceph_cap *cap;
1956 	int delayed = 0;
1957 
1958 	spin_lock(&ci->i_ceph_lock);
1959 	cap = ci->i_auth_cap;
1960 	dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
1961 	     ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
1962 	__ceph_flush_snaps(ci, &session, 1);
1963 	if (ci->i_flushing_caps) {
1964 		delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1965 				     __ceph_caps_used(ci),
1966 				     __ceph_caps_wanted(ci),
1967 				     cap->issued | cap->implemented,
1968 				     ci->i_flushing_caps, NULL);
1969 		if (delayed) {
1970 			spin_lock(&ci->i_ceph_lock);
1971 			__cap_delay_requeue(mdsc, ci);
1972 			spin_unlock(&ci->i_ceph_lock);
1973 		}
1974 	} else {
1975 		spin_unlock(&ci->i_ceph_lock);
1976 	}
1977 }
1978 
1979 
1980 /*
1981  * Take references to capabilities we hold, so that we don't release
1982  * them to the MDS prematurely.
1983  *
1984  * Protected by i_ceph_lock.
1985  */
__take_cap_refs(struct ceph_inode_info * ci,int got)1986 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1987 {
1988 	if (got & CEPH_CAP_PIN)
1989 		ci->i_pin_ref++;
1990 	if (got & CEPH_CAP_FILE_RD)
1991 		ci->i_rd_ref++;
1992 	if (got & CEPH_CAP_FILE_CACHE)
1993 		ci->i_rdcache_ref++;
1994 	if (got & CEPH_CAP_FILE_WR)
1995 		ci->i_wr_ref++;
1996 	if (got & CEPH_CAP_FILE_BUFFER) {
1997 		if (ci->i_wb_ref == 0)
1998 			ihold(&ci->vfs_inode);
1999 		ci->i_wb_ref++;
2000 		dout("__take_cap_refs %p wb %d -> %d (?)\n",
2001 		     &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2002 	}
2003 }
2004 
2005 /*
2006  * Try to grab cap references.  Specify those refs we @want, and the
2007  * minimal set we @need.  Also include the larger offset we are writing
2008  * to (when applicable), and check against max_size here as well.
2009  * Note that caller is responsible for ensuring max_size increases are
2010  * requested from the MDS.
2011  */
try_get_cap_refs(struct ceph_inode_info * ci,int need,int want,int * got,loff_t endoff,int * check_max,int * err)2012 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2013 			    int *got, loff_t endoff, int *check_max, int *err)
2014 {
2015 	struct inode *inode = &ci->vfs_inode;
2016 	int ret = 0;
2017 	int have, implemented;
2018 	int file_wanted;
2019 
2020 	dout("get_cap_refs %p need %s want %s\n", inode,
2021 	     ceph_cap_string(need), ceph_cap_string(want));
2022 	spin_lock(&ci->i_ceph_lock);
2023 
2024 	/* make sure file is actually open */
2025 	file_wanted = __ceph_caps_file_wanted(ci);
2026 	if ((file_wanted & need) == 0) {
2027 		dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2028 		     ceph_cap_string(need), ceph_cap_string(file_wanted));
2029 		*err = -EBADF;
2030 		ret = 1;
2031 		goto out;
2032 	}
2033 
2034 	if (need & CEPH_CAP_FILE_WR) {
2035 		if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2036 			dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2037 			     inode, endoff, ci->i_max_size);
2038 			if (endoff > ci->i_wanted_max_size) {
2039 				*check_max = 1;
2040 				ret = 1;
2041 			}
2042 			goto out;
2043 		}
2044 		/*
2045 		 * If a sync write is in progress, we must wait, so that we
2046 		 * can get a final snapshot value for size+mtime.
2047 		 */
2048 		if (__ceph_have_pending_cap_snap(ci)) {
2049 			dout("get_cap_refs %p cap_snap_pending\n", inode);
2050 			goto out;
2051 		}
2052 	}
2053 	have = __ceph_caps_issued(ci, &implemented);
2054 
2055 	/*
2056 	 * disallow writes while a truncate is pending
2057 	 */
2058 	if (ci->i_truncate_pending)
2059 		have &= ~CEPH_CAP_FILE_WR;
2060 
2061 	if ((have & need) == need) {
2062 		/*
2063 		 * Look at (implemented & ~have & not) so that we keep waiting
2064 		 * on transition from wanted -> needed caps.  This is needed
2065 		 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2066 		 * going before a prior buffered writeback happens.
2067 		 */
2068 		int not = want & ~(have & need);
2069 		int revoking = implemented & ~have;
2070 		dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2071 		     inode, ceph_cap_string(have), ceph_cap_string(not),
2072 		     ceph_cap_string(revoking));
2073 		if ((revoking & not) == 0) {
2074 			*got = need | (have & want);
2075 			__take_cap_refs(ci, *got);
2076 			ret = 1;
2077 		}
2078 	} else {
2079 		dout("get_cap_refs %p have %s needed %s\n", inode,
2080 		     ceph_cap_string(have), ceph_cap_string(need));
2081 	}
2082 out:
2083 	spin_unlock(&ci->i_ceph_lock);
2084 	dout("get_cap_refs %p ret %d got %s\n", inode,
2085 	     ret, ceph_cap_string(*got));
2086 	return ret;
2087 }
2088 
2089 /*
2090  * Check the offset we are writing up to against our current
2091  * max_size.  If necessary, tell the MDS we want to write to
2092  * a larger offset.
2093  */
check_max_size(struct inode * inode,loff_t endoff)2094 static void check_max_size(struct inode *inode, loff_t endoff)
2095 {
2096 	struct ceph_inode_info *ci = ceph_inode(inode);
2097 	int check = 0;
2098 
2099 	/* do we need to explicitly request a larger max_size? */
2100 	spin_lock(&ci->i_ceph_lock);
2101 	if ((endoff >= ci->i_max_size ||
2102 	     endoff > (inode->i_size << 1)) &&
2103 	    endoff > ci->i_wanted_max_size) {
2104 		dout("write %p at large endoff %llu, req max_size\n",
2105 		     inode, endoff);
2106 		ci->i_wanted_max_size = endoff;
2107 		check = 1;
2108 	}
2109 	spin_unlock(&ci->i_ceph_lock);
2110 	if (check)
2111 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2112 }
2113 
2114 /*
2115  * Wait for caps, and take cap references.  If we can't get a WR cap
2116  * due to a small max_size, make sure we check_max_size (and possibly
2117  * ask the mds) so we don't get hung up indefinitely.
2118  */
ceph_get_caps(struct ceph_inode_info * ci,int need,int want,int * got,loff_t endoff)2119 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2120 		  loff_t endoff)
2121 {
2122 	int check_max, ret, err;
2123 
2124 retry:
2125 	if (endoff > 0)
2126 		check_max_size(&ci->vfs_inode, endoff);
2127 	check_max = 0;
2128 	err = 0;
2129 	ret = wait_event_interruptible(ci->i_cap_wq,
2130 				       try_get_cap_refs(ci, need, want,
2131 							got, endoff,
2132 							&check_max, &err));
2133 	if (err)
2134 		ret = err;
2135 	if (check_max)
2136 		goto retry;
2137 	return ret;
2138 }
2139 
2140 /*
2141  * Take cap refs.  Caller must already know we hold at least one ref
2142  * on the caps in question or we don't know this is safe.
2143  */
ceph_get_cap_refs(struct ceph_inode_info * ci,int caps)2144 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2145 {
2146 	spin_lock(&ci->i_ceph_lock);
2147 	__take_cap_refs(ci, caps);
2148 	spin_unlock(&ci->i_ceph_lock);
2149 }
2150 
2151 /*
2152  * Release cap refs.
2153  *
2154  * If we released the last ref on any given cap, call ceph_check_caps
2155  * to release (or schedule a release).
2156  *
2157  * If we are releasing a WR cap (from a sync write), finalize any affected
2158  * cap_snap, and wake up any waiters.
2159  */
ceph_put_cap_refs(struct ceph_inode_info * ci,int had)2160 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2161 {
2162 	struct inode *inode = &ci->vfs_inode;
2163 	int last = 0, put = 0, flushsnaps = 0, wake = 0;
2164 	struct ceph_cap_snap *capsnap;
2165 
2166 	spin_lock(&ci->i_ceph_lock);
2167 	if (had & CEPH_CAP_PIN)
2168 		--ci->i_pin_ref;
2169 	if (had & CEPH_CAP_FILE_RD)
2170 		if (--ci->i_rd_ref == 0)
2171 			last++;
2172 	if (had & CEPH_CAP_FILE_CACHE)
2173 		if (--ci->i_rdcache_ref == 0)
2174 			last++;
2175 	if (had & CEPH_CAP_FILE_BUFFER) {
2176 		if (--ci->i_wb_ref == 0) {
2177 			last++;
2178 			put++;
2179 		}
2180 		dout("put_cap_refs %p wb %d -> %d (?)\n",
2181 		     inode, ci->i_wb_ref+1, ci->i_wb_ref);
2182 	}
2183 	if (had & CEPH_CAP_FILE_WR)
2184 		if (--ci->i_wr_ref == 0) {
2185 			last++;
2186 			if (!list_empty(&ci->i_cap_snaps)) {
2187 				capsnap = list_first_entry(&ci->i_cap_snaps,
2188 						     struct ceph_cap_snap,
2189 						     ci_item);
2190 				if (capsnap->writing) {
2191 					capsnap->writing = 0;
2192 					flushsnaps =
2193 						__ceph_finish_cap_snap(ci,
2194 								       capsnap);
2195 					wake = 1;
2196 				}
2197 			}
2198 		}
2199 	spin_unlock(&ci->i_ceph_lock);
2200 
2201 	dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2202 	     last ? " last" : "", put ? " put" : "");
2203 
2204 	if (last && !flushsnaps)
2205 		ceph_check_caps(ci, 0, NULL);
2206 	else if (flushsnaps)
2207 		ceph_flush_snaps(ci);
2208 	if (wake)
2209 		wake_up_all(&ci->i_cap_wq);
2210 	if (put)
2211 		iput(inode);
2212 }
2213 
2214 /*
2215  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2216  * context.  Adjust per-snap dirty page accounting as appropriate.
2217  * Once all dirty data for a cap_snap is flushed, flush snapped file
2218  * metadata back to the MDS.  If we dropped the last ref, call
2219  * ceph_check_caps.
2220  */
ceph_put_wrbuffer_cap_refs(struct ceph_inode_info * ci,int nr,struct ceph_snap_context * snapc)2221 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2222 				struct ceph_snap_context *snapc)
2223 {
2224 	struct inode *inode = &ci->vfs_inode;
2225 	int last = 0;
2226 	int complete_capsnap = 0;
2227 	int drop_capsnap = 0;
2228 	int found = 0;
2229 	struct ceph_cap_snap *capsnap = NULL;
2230 
2231 	spin_lock(&ci->i_ceph_lock);
2232 	ci->i_wrbuffer_ref -= nr;
2233 	last = !ci->i_wrbuffer_ref;
2234 
2235 	if (ci->i_head_snapc == snapc) {
2236 		ci->i_wrbuffer_ref_head -= nr;
2237 		if (ci->i_wrbuffer_ref_head == 0 &&
2238 		    ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2239 			BUG_ON(!ci->i_head_snapc);
2240 			ceph_put_snap_context(ci->i_head_snapc);
2241 			ci->i_head_snapc = NULL;
2242 		}
2243 		dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2244 		     inode,
2245 		     ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2246 		     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2247 		     last ? " LAST" : "");
2248 	} else {
2249 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2250 			if (capsnap->context == snapc) {
2251 				found = 1;
2252 				break;
2253 			}
2254 		}
2255 		BUG_ON(!found);
2256 		capsnap->dirty_pages -= nr;
2257 		if (capsnap->dirty_pages == 0) {
2258 			complete_capsnap = 1;
2259 			if (capsnap->dirty == 0)
2260 				/* cap writeback completed before we created
2261 				 * the cap_snap; no FLUSHSNAP is needed */
2262 				drop_capsnap = 1;
2263 		}
2264 		dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2265 		     " snap %lld %d/%d -> %d/%d %s%s%s\n",
2266 		     inode, capsnap, capsnap->context->seq,
2267 		     ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2268 		     ci->i_wrbuffer_ref, capsnap->dirty_pages,
2269 		     last ? " (wrbuffer last)" : "",
2270 		     complete_capsnap ? " (complete capsnap)" : "",
2271 		     drop_capsnap ? " (drop capsnap)" : "");
2272 		if (drop_capsnap) {
2273 			ceph_put_snap_context(capsnap->context);
2274 			list_del(&capsnap->ci_item);
2275 			list_del(&capsnap->flushing_item);
2276 			ceph_put_cap_snap(capsnap);
2277 		}
2278 	}
2279 
2280 	spin_unlock(&ci->i_ceph_lock);
2281 
2282 	if (last) {
2283 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2284 		iput(inode);
2285 	} else if (complete_capsnap) {
2286 		ceph_flush_snaps(ci);
2287 		wake_up_all(&ci->i_cap_wq);
2288 	}
2289 	if (drop_capsnap)
2290 		iput(inode);
2291 }
2292 
2293 /*
2294  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
2295  * actually be a revocation if it specifies a smaller cap set.)
2296  *
2297  * caller holds s_mutex and i_ceph_lock, we drop both.
2298  *
2299  * return value:
2300  *  0 - ok
2301  *  1 - check_caps on auth cap only (writeback)
2302  *  2 - check_caps (ack revoke)
2303  */
handle_cap_grant(struct inode * inode,struct ceph_mds_caps * grant,struct ceph_mds_session * session,struct ceph_cap * cap,struct ceph_buffer * xattr_buf)2304 static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2305 			     struct ceph_mds_session *session,
2306 			     struct ceph_cap *cap,
2307 			     struct ceph_buffer *xattr_buf)
2308 		__releases(ci->i_ceph_lock)
2309 {
2310 	struct ceph_inode_info *ci = ceph_inode(inode);
2311 	int mds = session->s_mds;
2312 	int seq = le32_to_cpu(grant->seq);
2313 	int newcaps = le32_to_cpu(grant->caps);
2314 	int issued, implemented, used, wanted, dirty;
2315 	u64 size = le64_to_cpu(grant->size);
2316 	u64 max_size = le64_to_cpu(grant->max_size);
2317 	struct timespec mtime, atime, ctime;
2318 	int check_caps = 0;
2319 	int wake = 0;
2320 	int writeback = 0;
2321 	int revoked_rdcache = 0;
2322 	int queue_invalidate = 0;
2323 
2324 	dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2325 	     inode, cap, mds, seq, ceph_cap_string(newcaps));
2326 	dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2327 		inode->i_size);
2328 
2329 	/*
2330 	 * If CACHE is being revoked, and we have no dirty buffers,
2331 	 * try to invalidate (once).  (If there are dirty buffers, we
2332 	 * will invalidate _after_ writeback.)
2333 	 */
2334 	if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2335 	    (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2336 	    !ci->i_wrbuffer_ref) {
2337 		if (try_nonblocking_invalidate(inode) == 0) {
2338 			revoked_rdcache = 1;
2339 		} else {
2340 			/* there were locked pages.. invalidate later
2341 			   in a separate thread. */
2342 			if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2343 				queue_invalidate = 1;
2344 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
2345 			}
2346 		}
2347 	}
2348 
2349 	/* side effects now are allowed */
2350 
2351 	issued = __ceph_caps_issued(ci, &implemented);
2352 	issued |= implemented | __ceph_caps_dirty(ci);
2353 
2354 	cap->cap_gen = session->s_cap_gen;
2355 
2356 	__check_cap_issue(ci, cap, newcaps);
2357 
2358 	if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2359 		inode->i_mode = le32_to_cpu(grant->mode);
2360 		inode->i_uid = le32_to_cpu(grant->uid);
2361 		inode->i_gid = le32_to_cpu(grant->gid);
2362 		dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2363 		     inode->i_uid, inode->i_gid);
2364 	}
2365 
2366 	if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2367 		set_nlink(inode, le32_to_cpu(grant->nlink));
2368 
2369 	if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2370 		int len = le32_to_cpu(grant->xattr_len);
2371 		u64 version = le64_to_cpu(grant->xattr_version);
2372 
2373 		if (version > ci->i_xattrs.version) {
2374 			dout(" got new xattrs v%llu on %p len %d\n",
2375 			     version, inode, len);
2376 			if (ci->i_xattrs.blob)
2377 				ceph_buffer_put(ci->i_xattrs.blob);
2378 			ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2379 			ci->i_xattrs.version = version;
2380 		}
2381 	}
2382 
2383 	/* size/ctime/mtime/atime? */
2384 	ceph_fill_file_size(inode, issued,
2385 			    le32_to_cpu(grant->truncate_seq),
2386 			    le64_to_cpu(grant->truncate_size), size);
2387 	ceph_decode_timespec(&mtime, &grant->mtime);
2388 	ceph_decode_timespec(&atime, &grant->atime);
2389 	ceph_decode_timespec(&ctime, &grant->ctime);
2390 	ceph_fill_file_time(inode, issued,
2391 			    le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2392 			    &atime);
2393 
2394 	/* max size increase? */
2395 	if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2396 		dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2397 		ci->i_max_size = max_size;
2398 		if (max_size >= ci->i_wanted_max_size) {
2399 			ci->i_wanted_max_size = 0;  /* reset */
2400 			ci->i_requested_max_size = 0;
2401 		}
2402 		wake = 1;
2403 	}
2404 
2405 	/* check cap bits */
2406 	wanted = __ceph_caps_wanted(ci);
2407 	used = __ceph_caps_used(ci);
2408 	dirty = __ceph_caps_dirty(ci);
2409 	dout(" my wanted = %s, used = %s, dirty %s\n",
2410 	     ceph_cap_string(wanted),
2411 	     ceph_cap_string(used),
2412 	     ceph_cap_string(dirty));
2413 	if (wanted != le32_to_cpu(grant->wanted)) {
2414 		dout("mds wanted %s -> %s\n",
2415 		     ceph_cap_string(le32_to_cpu(grant->wanted)),
2416 		     ceph_cap_string(wanted));
2417 		grant->wanted = cpu_to_le32(wanted);
2418 	}
2419 
2420 	cap->seq = seq;
2421 
2422 	/* file layout may have changed */
2423 	ci->i_layout = grant->layout;
2424 
2425 	/* revocation, grant, or no-op? */
2426 	if (cap->issued & ~newcaps) {
2427 		int revoking = cap->issued & ~newcaps;
2428 
2429 		dout("revocation: %s -> %s (revoking %s)\n",
2430 		     ceph_cap_string(cap->issued),
2431 		     ceph_cap_string(newcaps),
2432 		     ceph_cap_string(revoking));
2433 		if (revoking & used & CEPH_CAP_FILE_BUFFER)
2434 			writeback = 1;  /* initiate writeback; will delay ack */
2435 		else if (revoking == CEPH_CAP_FILE_CACHE &&
2436 			 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2437 			 queue_invalidate)
2438 			; /* do nothing yet, invalidation will be queued */
2439 		else if (cap == ci->i_auth_cap)
2440 			check_caps = 1; /* check auth cap only */
2441 		else
2442 			check_caps = 2; /* check all caps */
2443 		cap->issued = newcaps;
2444 		cap->implemented |= newcaps;
2445 	} else if (cap->issued == newcaps) {
2446 		dout("caps unchanged: %s -> %s\n",
2447 		     ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2448 	} else {
2449 		dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2450 		     ceph_cap_string(newcaps));
2451 		cap->issued = newcaps;
2452 		cap->implemented |= newcaps; /* add bits only, to
2453 					      * avoid stepping on a
2454 					      * pending revocation */
2455 		wake = 1;
2456 	}
2457 	BUG_ON(cap->issued & ~cap->implemented);
2458 
2459 	spin_unlock(&ci->i_ceph_lock);
2460 	if (writeback)
2461 		/*
2462 		 * queue inode for writeback: we can't actually call
2463 		 * filemap_write_and_wait, etc. from message handler
2464 		 * context.
2465 		 */
2466 		ceph_queue_writeback(inode);
2467 	if (queue_invalidate)
2468 		ceph_queue_invalidate(inode);
2469 	if (wake)
2470 		wake_up_all(&ci->i_cap_wq);
2471 
2472 	if (check_caps == 1)
2473 		ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2474 				session);
2475 	else if (check_caps == 2)
2476 		ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2477 	else
2478 		mutex_unlock(&session->s_mutex);
2479 }
2480 
2481 /*
2482  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2483  * MDS has been safely committed.
2484  */
handle_cap_flush_ack(struct inode * inode,u64 flush_tid,struct ceph_mds_caps * m,struct ceph_mds_session * session,struct ceph_cap * cap)2485 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2486 				 struct ceph_mds_caps *m,
2487 				 struct ceph_mds_session *session,
2488 				 struct ceph_cap *cap)
2489 	__releases(ci->i_ceph_lock)
2490 {
2491 	struct ceph_inode_info *ci = ceph_inode(inode);
2492 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2493 	unsigned seq = le32_to_cpu(m->seq);
2494 	int dirty = le32_to_cpu(m->dirty);
2495 	int cleaned = 0;
2496 	int drop = 0;
2497 	int i;
2498 
2499 	for (i = 0; i < CEPH_CAP_BITS; i++)
2500 		if ((dirty & (1 << i)) &&
2501 		    flush_tid == ci->i_cap_flush_tid[i])
2502 			cleaned |= 1 << i;
2503 
2504 	dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2505 	     " flushing %s -> %s\n",
2506 	     inode, session->s_mds, seq, ceph_cap_string(dirty),
2507 	     ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2508 	     ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2509 
2510 	if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2511 		goto out;
2512 
2513 	ci->i_flushing_caps &= ~cleaned;
2514 
2515 	spin_lock(&mdsc->cap_dirty_lock);
2516 	if (ci->i_flushing_caps == 0) {
2517 		list_del_init(&ci->i_flushing_item);
2518 		if (!list_empty(&session->s_cap_flushing))
2519 			dout(" mds%d still flushing cap on %p\n",
2520 			     session->s_mds,
2521 			     &list_entry(session->s_cap_flushing.next,
2522 					 struct ceph_inode_info,
2523 					 i_flushing_item)->vfs_inode);
2524 		mdsc->num_cap_flushing--;
2525 		wake_up_all(&mdsc->cap_flushing_wq);
2526 		dout(" inode %p now !flushing\n", inode);
2527 
2528 		if (ci->i_dirty_caps == 0) {
2529 			dout(" inode %p now clean\n", inode);
2530 			BUG_ON(!list_empty(&ci->i_dirty_item));
2531 			drop = 1;
2532 			if (ci->i_wrbuffer_ref_head == 0) {
2533 				BUG_ON(!ci->i_head_snapc);
2534 				ceph_put_snap_context(ci->i_head_snapc);
2535 				ci->i_head_snapc = NULL;
2536 			}
2537 		} else {
2538 			BUG_ON(list_empty(&ci->i_dirty_item));
2539 		}
2540 	}
2541 	spin_unlock(&mdsc->cap_dirty_lock);
2542 	wake_up_all(&ci->i_cap_wq);
2543 
2544 out:
2545 	spin_unlock(&ci->i_ceph_lock);
2546 	if (drop)
2547 		iput(inode);
2548 }
2549 
2550 /*
2551  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
2552  * throw away our cap_snap.
2553  *
2554  * Caller hold s_mutex.
2555  */
handle_cap_flushsnap_ack(struct inode * inode,u64 flush_tid,struct ceph_mds_caps * m,struct ceph_mds_session * session)2556 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2557 				     struct ceph_mds_caps *m,
2558 				     struct ceph_mds_session *session)
2559 {
2560 	struct ceph_inode_info *ci = ceph_inode(inode);
2561 	u64 follows = le64_to_cpu(m->snap_follows);
2562 	struct ceph_cap_snap *capsnap;
2563 	int drop = 0;
2564 
2565 	dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2566 	     inode, ci, session->s_mds, follows);
2567 
2568 	spin_lock(&ci->i_ceph_lock);
2569 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2570 		if (capsnap->follows == follows) {
2571 			if (capsnap->flush_tid != flush_tid) {
2572 				dout(" cap_snap %p follows %lld tid %lld !="
2573 				     " %lld\n", capsnap, follows,
2574 				     flush_tid, capsnap->flush_tid);
2575 				break;
2576 			}
2577 			WARN_ON(capsnap->dirty_pages || capsnap->writing);
2578 			dout(" removing %p cap_snap %p follows %lld\n",
2579 			     inode, capsnap, follows);
2580 			ceph_put_snap_context(capsnap->context);
2581 			list_del(&capsnap->ci_item);
2582 			list_del(&capsnap->flushing_item);
2583 			ceph_put_cap_snap(capsnap);
2584 			drop = 1;
2585 			break;
2586 		} else {
2587 			dout(" skipping cap_snap %p follows %lld\n",
2588 			     capsnap, capsnap->follows);
2589 		}
2590 	}
2591 	spin_unlock(&ci->i_ceph_lock);
2592 	if (drop)
2593 		iput(inode);
2594 }
2595 
2596 /*
2597  * Handle TRUNC from MDS, indicating file truncation.
2598  *
2599  * caller hold s_mutex.
2600  */
handle_cap_trunc(struct inode * inode,struct ceph_mds_caps * trunc,struct ceph_mds_session * session)2601 static void handle_cap_trunc(struct inode *inode,
2602 			     struct ceph_mds_caps *trunc,
2603 			     struct ceph_mds_session *session)
2604 	__releases(ci->i_ceph_lock)
2605 {
2606 	struct ceph_inode_info *ci = ceph_inode(inode);
2607 	int mds = session->s_mds;
2608 	int seq = le32_to_cpu(trunc->seq);
2609 	u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2610 	u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2611 	u64 size = le64_to_cpu(trunc->size);
2612 	int implemented = 0;
2613 	int dirty = __ceph_caps_dirty(ci);
2614 	int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2615 	int queue_trunc = 0;
2616 
2617 	issued |= implemented | dirty;
2618 
2619 	dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2620 	     inode, mds, seq, truncate_size, truncate_seq);
2621 	queue_trunc = ceph_fill_file_size(inode, issued,
2622 					  truncate_seq, truncate_size, size);
2623 	spin_unlock(&ci->i_ceph_lock);
2624 
2625 	if (queue_trunc)
2626 		ceph_queue_vmtruncate(inode);
2627 }
2628 
2629 /*
2630  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
2631  * different one.  If we are the most recent migration we've seen (as
2632  * indicated by mseq), make note of the migrating cap bits for the
2633  * duration (until we see the corresponding IMPORT).
2634  *
2635  * caller holds s_mutex
2636  */
handle_cap_export(struct inode * inode,struct ceph_mds_caps * ex,struct ceph_mds_session * session,int * open_target_sessions)2637 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2638 			      struct ceph_mds_session *session,
2639 			      int *open_target_sessions)
2640 {
2641 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2642 	struct ceph_inode_info *ci = ceph_inode(inode);
2643 	int mds = session->s_mds;
2644 	unsigned mseq = le32_to_cpu(ex->migrate_seq);
2645 	struct ceph_cap *cap = NULL, *t;
2646 	struct rb_node *p;
2647 	int remember = 1;
2648 
2649 	dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2650 	     inode, ci, mds, mseq);
2651 
2652 	spin_lock(&ci->i_ceph_lock);
2653 
2654 	/* make sure we haven't seen a higher mseq */
2655 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2656 		t = rb_entry(p, struct ceph_cap, ci_node);
2657 		if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2658 			dout(" higher mseq on cap from mds%d\n",
2659 			     t->session->s_mds);
2660 			remember = 0;
2661 		}
2662 		if (t->session->s_mds == mds)
2663 			cap = t;
2664 	}
2665 
2666 	if (cap) {
2667 		if (remember) {
2668 			/* make note */
2669 			ci->i_cap_exporting_mds = mds;
2670 			ci->i_cap_exporting_mseq = mseq;
2671 			ci->i_cap_exporting_issued = cap->issued;
2672 
2673 			/*
2674 			 * make sure we have open sessions with all possible
2675 			 * export targets, so that we get the matching IMPORT
2676 			 */
2677 			*open_target_sessions = 1;
2678 
2679 			/*
2680 			 * we can't flush dirty caps that we've seen the
2681 			 * EXPORT but no IMPORT for
2682 			 */
2683 			spin_lock(&mdsc->cap_dirty_lock);
2684 			if (!list_empty(&ci->i_dirty_item)) {
2685 				dout(" moving %p to cap_dirty_migrating\n",
2686 				     inode);
2687 				list_move(&ci->i_dirty_item,
2688 					  &mdsc->cap_dirty_migrating);
2689 			}
2690 			spin_unlock(&mdsc->cap_dirty_lock);
2691 		}
2692 		__ceph_remove_cap(cap);
2693 	}
2694 	/* else, we already released it */
2695 
2696 	spin_unlock(&ci->i_ceph_lock);
2697 }
2698 
2699 /*
2700  * Handle cap IMPORT.  If there are temp bits from an older EXPORT,
2701  * clean them up.
2702  *
2703  * caller holds s_mutex.
2704  */
handle_cap_import(struct ceph_mds_client * mdsc,struct inode * inode,struct ceph_mds_caps * im,struct ceph_mds_session * session,void * snaptrace,int snaptrace_len)2705 static void handle_cap_import(struct ceph_mds_client *mdsc,
2706 			      struct inode *inode, struct ceph_mds_caps *im,
2707 			      struct ceph_mds_session *session,
2708 			      void *snaptrace, int snaptrace_len)
2709 {
2710 	struct ceph_inode_info *ci = ceph_inode(inode);
2711 	int mds = session->s_mds;
2712 	unsigned issued = le32_to_cpu(im->caps);
2713 	unsigned wanted = le32_to_cpu(im->wanted);
2714 	unsigned seq = le32_to_cpu(im->seq);
2715 	unsigned mseq = le32_to_cpu(im->migrate_seq);
2716 	u64 realmino = le64_to_cpu(im->realm);
2717 	u64 cap_id = le64_to_cpu(im->cap_id);
2718 
2719 	if (ci->i_cap_exporting_mds >= 0 &&
2720 	    ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2721 		dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2722 		     " - cleared exporting from mds%d\n",
2723 		     inode, ci, mds, mseq,
2724 		     ci->i_cap_exporting_mds);
2725 		ci->i_cap_exporting_issued = 0;
2726 		ci->i_cap_exporting_mseq = 0;
2727 		ci->i_cap_exporting_mds = -1;
2728 
2729 		spin_lock(&mdsc->cap_dirty_lock);
2730 		if (!list_empty(&ci->i_dirty_item)) {
2731 			dout(" moving %p back to cap_dirty\n", inode);
2732 			list_move(&ci->i_dirty_item, &mdsc->cap_dirty);
2733 		}
2734 		spin_unlock(&mdsc->cap_dirty_lock);
2735 	} else {
2736 		dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2737 		     inode, ci, mds, mseq);
2738 	}
2739 
2740 	down_write(&mdsc->snap_rwsem);
2741 	ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2742 			       false);
2743 	downgrade_write(&mdsc->snap_rwsem);
2744 	ceph_add_cap(inode, session, cap_id, -1,
2745 		     issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2746 		     NULL /* no caps context */);
2747 	kick_flushing_inode_caps(mdsc, session, inode);
2748 	up_read(&mdsc->snap_rwsem);
2749 
2750 	/* make sure we re-request max_size, if necessary */
2751 	spin_lock(&ci->i_ceph_lock);
2752 	ci->i_wanted_max_size = 0;  /* reset */
2753 	ci->i_requested_max_size = 0;
2754 	spin_unlock(&ci->i_ceph_lock);
2755 }
2756 
2757 /*
2758  * Handle a caps message from the MDS.
2759  *
2760  * Identify the appropriate session, inode, and call the right handler
2761  * based on the cap op.
2762  */
ceph_handle_caps(struct ceph_mds_session * session,struct ceph_msg * msg)2763 void ceph_handle_caps(struct ceph_mds_session *session,
2764 		      struct ceph_msg *msg)
2765 {
2766 	struct ceph_mds_client *mdsc = session->s_mdsc;
2767 	struct super_block *sb = mdsc->fsc->sb;
2768 	struct inode *inode;
2769 	struct ceph_inode_info *ci;
2770 	struct ceph_cap *cap;
2771 	struct ceph_mds_caps *h;
2772 	int mds = session->s_mds;
2773 	int op;
2774 	u32 seq, mseq;
2775 	struct ceph_vino vino;
2776 	u64 cap_id;
2777 	u64 size, max_size;
2778 	u64 tid;
2779 	void *snaptrace;
2780 	size_t snaptrace_len;
2781 	void *flock;
2782 	u32 flock_len;
2783 	int open_target_sessions = 0;
2784 
2785 	dout("handle_caps from mds%d\n", mds);
2786 
2787 	/* decode */
2788 	tid = le64_to_cpu(msg->hdr.tid);
2789 	if (msg->front.iov_len < sizeof(*h))
2790 		goto bad;
2791 	h = msg->front.iov_base;
2792 	op = le32_to_cpu(h->op);
2793 	vino.ino = le64_to_cpu(h->ino);
2794 	vino.snap = CEPH_NOSNAP;
2795 	cap_id = le64_to_cpu(h->cap_id);
2796 	seq = le32_to_cpu(h->seq);
2797 	mseq = le32_to_cpu(h->migrate_seq);
2798 	size = le64_to_cpu(h->size);
2799 	max_size = le64_to_cpu(h->max_size);
2800 
2801 	snaptrace = h + 1;
2802 	snaptrace_len = le32_to_cpu(h->snap_trace_len);
2803 
2804 	if (le16_to_cpu(msg->hdr.version) >= 2) {
2805 		void *p, *end;
2806 
2807 		p = snaptrace + snaptrace_len;
2808 		end = msg->front.iov_base + msg->front.iov_len;
2809 		ceph_decode_32_safe(&p, end, flock_len, bad);
2810 		flock = p;
2811 	} else {
2812 		flock = NULL;
2813 		flock_len = 0;
2814 	}
2815 
2816 	mutex_lock(&session->s_mutex);
2817 	session->s_seq++;
2818 	dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2819 	     (unsigned)seq);
2820 
2821 	/* lookup ino */
2822 	inode = ceph_find_inode(sb, vino);
2823 	ci = ceph_inode(inode);
2824 	dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2825 	     vino.snap, inode);
2826 	if (!inode) {
2827 		dout(" i don't have ino %llx\n", vino.ino);
2828 
2829 		if (op == CEPH_CAP_OP_IMPORT)
2830 			__queue_cap_release(session, vino.ino, cap_id,
2831 					    mseq, seq);
2832 		goto flush_cap_releases;
2833 	}
2834 
2835 	/* these will work even if we don't have a cap yet */
2836 	switch (op) {
2837 	case CEPH_CAP_OP_FLUSHSNAP_ACK:
2838 		handle_cap_flushsnap_ack(inode, tid, h, session);
2839 		goto done;
2840 
2841 	case CEPH_CAP_OP_EXPORT:
2842 		handle_cap_export(inode, h, session, &open_target_sessions);
2843 		goto done;
2844 
2845 	case CEPH_CAP_OP_IMPORT:
2846 		handle_cap_import(mdsc, inode, h, session,
2847 				  snaptrace, snaptrace_len);
2848 	}
2849 
2850 	/* the rest require a cap */
2851 	spin_lock(&ci->i_ceph_lock);
2852 	cap = __get_cap_for_mds(ceph_inode(inode), mds);
2853 	if (!cap) {
2854 		dout(" no cap on %p ino %llx.%llx from mds%d\n",
2855 		     inode, ceph_ino(inode), ceph_snap(inode), mds);
2856 		spin_unlock(&ci->i_ceph_lock);
2857 		goto flush_cap_releases;
2858 	}
2859 
2860 	/* note that each of these drops i_ceph_lock for us */
2861 	switch (op) {
2862 	case CEPH_CAP_OP_REVOKE:
2863 	case CEPH_CAP_OP_GRANT:
2864 	case CEPH_CAP_OP_IMPORT:
2865 		handle_cap_grant(inode, h, session, cap, msg->middle);
2866 		goto done_unlocked;
2867 
2868 	case CEPH_CAP_OP_FLUSH_ACK:
2869 		handle_cap_flush_ack(inode, tid, h, session, cap);
2870 		break;
2871 
2872 	case CEPH_CAP_OP_TRUNC:
2873 		handle_cap_trunc(inode, h, session);
2874 		break;
2875 
2876 	default:
2877 		spin_unlock(&ci->i_ceph_lock);
2878 		pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2879 		       ceph_cap_op_name(op));
2880 	}
2881 
2882 	goto done;
2883 
2884 flush_cap_releases:
2885 	/*
2886 	 * send any full release message to try to move things
2887 	 * along for the mds (who clearly thinks we still have this
2888 	 * cap).
2889 	 */
2890 	ceph_add_cap_releases(mdsc, session);
2891 	ceph_send_cap_releases(mdsc, session);
2892 
2893 done:
2894 	mutex_unlock(&session->s_mutex);
2895 done_unlocked:
2896 	if (inode)
2897 		iput(inode);
2898 	if (open_target_sessions)
2899 		ceph_mdsc_open_export_target_sessions(mdsc, session);
2900 	return;
2901 
2902 bad:
2903 	pr_err("ceph_handle_caps: corrupt message\n");
2904 	ceph_msg_dump(msg);
2905 	return;
2906 }
2907 
2908 /*
2909  * Delayed work handler to process end of delayed cap release LRU list.
2910  */
ceph_check_delayed_caps(struct ceph_mds_client * mdsc)2911 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2912 {
2913 	struct ceph_inode_info *ci;
2914 	int flags = CHECK_CAPS_NODELAY;
2915 
2916 	dout("check_delayed_caps\n");
2917 	while (1) {
2918 		spin_lock(&mdsc->cap_delay_lock);
2919 		if (list_empty(&mdsc->cap_delay_list))
2920 			break;
2921 		ci = list_first_entry(&mdsc->cap_delay_list,
2922 				      struct ceph_inode_info,
2923 				      i_cap_delay_list);
2924 		if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2925 		    time_before(jiffies, ci->i_hold_caps_max))
2926 			break;
2927 		list_del_init(&ci->i_cap_delay_list);
2928 		spin_unlock(&mdsc->cap_delay_lock);
2929 		dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2930 		ceph_check_caps(ci, flags, NULL);
2931 	}
2932 	spin_unlock(&mdsc->cap_delay_lock);
2933 }
2934 
2935 /*
2936  * Flush all dirty caps to the mds
2937  */
ceph_flush_dirty_caps(struct ceph_mds_client * mdsc)2938 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2939 {
2940 	struct ceph_inode_info *ci;
2941 	struct inode *inode;
2942 
2943 	dout("flush_dirty_caps\n");
2944 	spin_lock(&mdsc->cap_dirty_lock);
2945 	while (!list_empty(&mdsc->cap_dirty)) {
2946 		ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
2947 				      i_dirty_item);
2948 		inode = &ci->vfs_inode;
2949 		ihold(inode);
2950 		dout("flush_dirty_caps %p\n", inode);
2951 		spin_unlock(&mdsc->cap_dirty_lock);
2952 		ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
2953 		iput(inode);
2954 		spin_lock(&mdsc->cap_dirty_lock);
2955 	}
2956 	spin_unlock(&mdsc->cap_dirty_lock);
2957 	dout("flush_dirty_caps done\n");
2958 }
2959 
2960 /*
2961  * Drop open file reference.  If we were the last open file,
2962  * we may need to release capabilities to the MDS (or schedule
2963  * their delayed release).
2964  */
ceph_put_fmode(struct ceph_inode_info * ci,int fmode)2965 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2966 {
2967 	struct inode *inode = &ci->vfs_inode;
2968 	int last = 0;
2969 
2970 	spin_lock(&ci->i_ceph_lock);
2971 	dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2972 	     ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2973 	BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2974 	if (--ci->i_nr_by_mode[fmode] == 0)
2975 		last++;
2976 	spin_unlock(&ci->i_ceph_lock);
2977 
2978 	if (last && ci->i_vino.snap == CEPH_NOSNAP)
2979 		ceph_check_caps(ci, 0, NULL);
2980 }
2981 
2982 /*
2983  * Helpers for embedding cap and dentry lease releases into mds
2984  * requests.
2985  *
2986  * @force is used by dentry_release (below) to force inclusion of a
2987  * record for the directory inode, even when there aren't any caps to
2988  * drop.
2989  */
ceph_encode_inode_release(void ** p,struct inode * inode,int mds,int drop,int unless,int force)2990 int ceph_encode_inode_release(void **p, struct inode *inode,
2991 			      int mds, int drop, int unless, int force)
2992 {
2993 	struct ceph_inode_info *ci = ceph_inode(inode);
2994 	struct ceph_cap *cap;
2995 	struct ceph_mds_request_release *rel = *p;
2996 	int used, dirty;
2997 	int ret = 0;
2998 
2999 	spin_lock(&ci->i_ceph_lock);
3000 	used = __ceph_caps_used(ci);
3001 	dirty = __ceph_caps_dirty(ci);
3002 
3003 	dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3004 	     inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3005 	     ceph_cap_string(unless));
3006 
3007 	/* only drop unused, clean caps */
3008 	drop &= ~(used | dirty);
3009 
3010 	cap = __get_cap_for_mds(ci, mds);
3011 	if (cap && __cap_is_valid(cap)) {
3012 		if (force ||
3013 		    ((cap->issued & drop) &&
3014 		     (cap->issued & unless) == 0)) {
3015 			if ((cap->issued & drop) &&
3016 			    (cap->issued & unless) == 0) {
3017 				dout("encode_inode_release %p cap %p %s -> "
3018 				     "%s\n", inode, cap,
3019 				     ceph_cap_string(cap->issued),
3020 				     ceph_cap_string(cap->issued & ~drop));
3021 				cap->issued &= ~drop;
3022 				cap->implemented &= ~drop;
3023 				if (ci->i_ceph_flags & CEPH_I_NODELAY) {
3024 					int wanted = __ceph_caps_wanted(ci);
3025 					dout("  wanted %s -> %s (act %s)\n",
3026 					     ceph_cap_string(cap->mds_wanted),
3027 					     ceph_cap_string(cap->mds_wanted &
3028 							     ~wanted),
3029 					     ceph_cap_string(wanted));
3030 					cap->mds_wanted &= wanted;
3031 				}
3032 			} else {
3033 				dout("encode_inode_release %p cap %p %s"
3034 				     " (force)\n", inode, cap,
3035 				     ceph_cap_string(cap->issued));
3036 			}
3037 
3038 			rel->ino = cpu_to_le64(ceph_ino(inode));
3039 			rel->cap_id = cpu_to_le64(cap->cap_id);
3040 			rel->seq = cpu_to_le32(cap->seq);
3041 			rel->issue_seq = cpu_to_le32(cap->issue_seq),
3042 			rel->mseq = cpu_to_le32(cap->mseq);
3043 			rel->caps = cpu_to_le32(cap->issued);
3044 			rel->wanted = cpu_to_le32(cap->mds_wanted);
3045 			rel->dname_len = 0;
3046 			rel->dname_seq = 0;
3047 			*p += sizeof(*rel);
3048 			ret = 1;
3049 		} else {
3050 			dout("encode_inode_release %p cap %p %s\n",
3051 			     inode, cap, ceph_cap_string(cap->issued));
3052 		}
3053 	}
3054 	spin_unlock(&ci->i_ceph_lock);
3055 	return ret;
3056 }
3057 
ceph_encode_dentry_release(void ** p,struct dentry * dentry,int mds,int drop,int unless)3058 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3059 			       int mds, int drop, int unless)
3060 {
3061 	struct inode *dir = dentry->d_parent->d_inode;
3062 	struct ceph_mds_request_release *rel = *p;
3063 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3064 	int force = 0;
3065 	int ret;
3066 
3067 	/*
3068 	 * force an record for the directory caps if we have a dentry lease.
3069 	 * this is racy (can't take i_ceph_lock and d_lock together), but it
3070 	 * doesn't have to be perfect; the mds will revoke anything we don't
3071 	 * release.
3072 	 */
3073 	spin_lock(&dentry->d_lock);
3074 	if (di->lease_session && di->lease_session->s_mds == mds)
3075 		force = 1;
3076 	spin_unlock(&dentry->d_lock);
3077 
3078 	ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3079 
3080 	spin_lock(&dentry->d_lock);
3081 	if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3082 		dout("encode_dentry_release %p mds%d seq %d\n",
3083 		     dentry, mds, (int)di->lease_seq);
3084 		rel->dname_len = cpu_to_le32(dentry->d_name.len);
3085 		memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3086 		*p += dentry->d_name.len;
3087 		rel->dname_seq = cpu_to_le32(di->lease_seq);
3088 		__ceph_mdsc_drop_dentry_lease(dentry);
3089 	}
3090 	spin_unlock(&dentry->d_lock);
3091 	return ret;
3092 }
3093